arxiv_id
stringlengths 12
12
| markdown
stringlengths 20.2k
18.9M
| paper_doi
stringclasses 42
values | paper_authors
sequencelengths 1
198
| paper_published_date
stringdate 2024-07-07 06:39:16
2024-08-01 08:57:47
| paper_updated_date
stringdate 2024-07-07 06:39:16
2025-04-23 08:12:32
| categories
sequencelengths 1
5
| title
stringlengths 22
182
| summary
stringlengths 190
1.92k
|
|---|---|---|---|---|---|---|---|---|
2408.00385v1 | ## Quantitative Group Testing and Pooled Data in the Linear Regime with Sublinear Tests
Nelvin Tan, Pablo Pascual Cobo, and Ramji Venkataramanan
## Abstract
In the pooled data problem, the goal is to identify the categories associated with a large collection of items via a sequence of pooled tests. Each pooled test reveals the number of items in the pool belonging to each category. A prominent special case is quantitative group testing (QGT), which is the case of pooled data with two categories. We consider these problems in the non-adaptive and linear regime, where the fraction of items in each category is of constant order. We propose a scheme with a spatially coupled Bernoulli test matrix and an efficient approximate message passing (AMP) algorithm for recovery. We rigorously characterize its asymptotic performance in both the noiseless and noisy settings, and prove that in the noiseless case, the AMP algorithm achieves almost-exact recovery with a number of tests sublinear in the number of items. For both QGT and pooled data, this is the first efficient scheme that provably achieves recovery in the linear regime with a sublinear number of tests, with performance degrading gracefully in the presence of noise. Numerical simulations illustrate the benefits of the spatially coupled scheme at finite dimensions, showing that it outperforms i.i.d. test designs as well as other recovery algorithms based on convex programming.
## 1 Introduction
Consider a large collection of items, each of which is either defective or non-defective. In group testing [1], the goal is to identify the defective set via pooled tests, where groups of items are tested together, with as few tests as possible. In the original Boolean group testing model, which has been studied extensively [1-5], each test returns a positive outcome if it includes at least one defective item and a negative outcome otherwise. Its variant, the quantitative group testing (QGT) model [6], is useful when tests are more informative: each test reveals the number of defective items in that pool. QGT is of interest in a range of modern applications, including genomics [7], multi-access communication [8], and network traffic monitoring [9]. A more general version of QGT, where each item belongs to one of L > 2 categories, is known as the pooled data problem [10]. The goal is to identify the categories via a sequence of pooled tests, where each pooled test reveals the number of items of each category within the pool.
In this paper, we consider non-adaptive QGT and pooled data, where the tests are all designed in advance, making them amenable to being implemented in parallel. We also consider the linear regime, which for QGT, means that the number of defective items is proportional to the total number
N. Tan was supported by a Cambridge Trust Scholarship and the Harding Distinguished Postgraduate Scholars Programme Leverage Scheme. P. Pascual Cobo was supported by a Engineering and Physical Sciences Research Council Doctoral Training Award. This work was presented in part at the 2024 International Zurich Seminar on Information and Communication (IZS). The authors are with the Department of Engineering, University of Cambridge, UK. Emails: [email protected], [email protected], [email protected] .
of items. For pooled data, the linear regime implies that the proportion of items in each category is non-vanishing as the number of items increases, a realistic assumption in practical applications.
## 1.1 Problem Setup
Quantitative group testing. There are p items, whose status is denoted by the binary vector β ∈ { 0 1 , } p , where 1 represents a defective item and 0 a non-defective item. Items are allocated to tests using a binary design (or test) matrix X ∈ { 0 1 , } n × p , where n is the number of tests and p is the number of items. The i th row X i, : determines the pooling design of the i th test, where X ij = 1 indicates that the j th item will be included in the i th test, and X ij = 0 indicates otherwise. Let d be the number of defective items with d < p . We consider the linear regime, where the fraction of defective items d/p converges to π ∈ (0 1) , . Mathematically, we define the QGT model as
$$y _ { i } = \left ( X _ { i, \colon } \right ) ^ { \top } \beta + \Psi _ { i } \quad \text{for $i\in\{1,\dots,n\}$},$$
where y i is the i th element of y ∈ R n , X i, : is the i th row of X ∈ R n × p represented as a column vector, and Ψ i is the i th element of the additive noise Ψ ∈ R n . Under the noiseless setting (i.e., all entries of Ψ are zero), the output y i is the number of defective items in the i th test. The goal of QGT is to recover β with as few tests as possible.
We will use the almost-exact recovery criterion, which is achieved by an estimator ˜ β if
̸
$$\frac { 1 } { p } \sum _ { j = 1 } ^ { p } 1 \{ \tilde { \beta } _ { j } \neq \beta _ { j } \} \to 0 \text{ as } p \to \infty.$$
̸
This is a weaker notion of recovery as compared to the exact recovery criterion [11] where we want the probability of error P [ ˜ = β β ] → 0 as p →∞ . We note that an almost-exact recovery criterion is meaningful in the linear regime, but not in the sublinear regime where the number of defectives d = ( ) o p , since setting ˜ β to be the zero vector would trivially satisfy (2).
Pooled data. The signal to be estimated is a matrix B ∈ { 0 1 , } p × L , where each row is a onehot vector. For example B j, : = [0 1 0 , , , . . . , 0] represents the j th item belonging to category 2 (the position of one in B j, : ). We consider the linear regime, where the fraction of items in each category l converges to π l ∈ (0 1) , , where ∑ L l =1 π l = 1 . The model is
$$Y _ { i, \colon } = B ^ { \top } X _ { i, \colon } + \Psi _ { i, \colon } \in \mathbb { R } ^ { L } \quad \text{for $i\in\{1,\dots,n\}$},$$
where Y i, : is the i th row of Y ∈ R n × L represented as a column vector, and Ψ i, : is the i th row of the additive noise Ψ ∈ R n × L represented as a column vector. Under the noiseless setting (i.e., all entries of Ψ are zero), the output of each test Y i, : tells us the number of items from each category present in the test, which can be viewed as a histogram. Similar to QGT, an estimator ˜ B achieves almost-exact recovery if
̸
$$\frac { 1 } { p L } \sum _ { j = 1 } ^ { p } \sum _ { l = 1 } ^ { L } { \mathbf 1 } \{ \widetilde { B } _ { j l } \neq B _ { j l } \} \to 0 \text{ as } p \to \infty.$$
The number of categories L does not grow with p .
Information-theoretic limits. For noiseless pooled data in the linear regime, the informationtheoretic limit on the number of tests required was established by Scarlett and Cevher [11], closing the gap between previously derived upper and lower bounds [10,12,13]. It was shown in [11] that the minimum number of tests needed for exact recovery is n ∗ = γ ∗ p log p (1 + o (1)) , where
$$\gamma ^ { * } = \max _ { r \in \{ 1, \dots, L - 1 \} } \frac { 2 [ H ( \pi ) - H ( \pi ^ { ( r ) } ) ] } { L - r },$$
while H π ( ) = -∑ L l =1 π l log π l is the Shannon entropy function, and π ( r ) = ( π ( r ) 1 , . . . , π ( r ) r ) is a vector whose first entry sums the largest ( L -r + 1) entries of π , and whose remaining entries coincide with the remaining ( r -1) entries of π . Setting L = 2 above gives the information-theoretic limit for noiseless QGT.
For noisy pooled data where the entries of the noise matrix are independent with zero mean, it was shown in [11] that we require n = Ω( p log p ) tests for exact recovery, in contrast with the sublinear Θ ( p log p ) required in the noiseless case.
Efficient Algorithms. For pooled data, Wang et al. [10] proposed a deterministic design matrix and a polynomial-time algorithm that achieves exact recovery with n = Ω ( p log p ) tests, matching the optimal sample complexity above. However, both the test design and the recovery algorithm (based on Gaussian elimination) are tailored to the noiseless setting, and do not extend to the noisy case. In this paper, we focus on random test designs, which are more robust with respect to the items included in each test, and on recovery algorithms whose performance degrades gracefully with the noise level. For random i.i.d. designs (where X ij i . i . d . ∼ Bernoulli ( α ) for some α ∈ (0 1) , ), efficient recovery using Approximate Message Passing (AMP) algorithms was studied in [13, 14]. Rigorous guarantees on the recovery performance of AMP, established in [14], imply that with the i.i.d. Bernoulli design, the AMP algorithm needs n = Θ( ) p tests for almost-exact recovery, even in the noiseless QGT/pooled data setting. Thus, for random designs there is an orderlog p gap between the information-theoretic limit and the best known algorithms.
## 1.2 Approximate Message Passing and Spatial Coupling
Approximate message passing (AMP) is a family of iterative algorithms that can be tailored to take advantage of structural information about the signals and the model, e.g., a known prior on the signal vector or on the proportion of observations that come from each signal. AMP algorithms were first proposed for the standard linear model [15-18], but have since been applied to a range of statistical problems, including estimation in generalized linear models and their variants [19-23], as well as low-rank matrix and tensor estimation [24-30]. In all these settings, under suitable model assumptions the performance of AMP in the high-dimensional limit is characterized by a succinct deterministic recursion called state evolution . The literature on AMP is vast, and we refer the interested reader to [31] for a survey.
In this paper, we will use a spatially coupled design matrix X and a suitable AMP algorithm for recovery. The spatially coupled matrix has a block-wise structure, with blocks along a band-diagonal having i.i.d. Bernoulli entries and the remaining blocks being all zeros (see Figure 1). Our scheme is inspired by a line of work on compressed sensing with spatially coupled designs [18, 32-34]. For a noiseless linear model defined via a spatially coupled Gaussian sensing matrix, Donoho et al. [34] showed that AMP recovers the signal with high probability when the sampling ratio δ = n/p
exceeds the Rényi information dimension of the signal prior. The Rényi information dimension is zero for priors supported on a finite set, which implies that AMP can recover the signal with n = o p ( ) measurements. Recently, it was also shown that using a spatially coupled sensing matrix in a generalized linear model allows AMP to achieve the Bayes-optimal error (corresponding to an i.i.d. Gaussian matrix) [35].
## 1.3 Main Contributions
In Section 3, we describe the spatially coupled random test design and an AMP algorithm (SCAMP) for signal recovery. In Theorem 3.2 we give a precise characterization of the performance of SC-AMP in the asymptotic regime where the number of tests n grows proportionally with the number of items p (with n/p → δ , a constant). Using this characterization, we bound the MSE of the algorithm in the low noise regime (Theorem 3.7) and show that for noiseless QGT, the SCAMP algorithm achieves almost-exact recovery with probability one, for any sampling ratio δ > 0 (Corollary 3.8). This implies that it achieves almost-exact recovery with n = ( ) o p tests. In Section 4, we generalize the SC-AMP algorithm to the pooled data setting and again establish almost-exact recovery for any δ > 0 (Theorem 4.1).
At the heart of our theoretical guarantees is a rigorous analysis of an AMP algorithm for a generalized linear model (GLM) with a generic spatially coupled design matrix. (The matrix consists of blocks of independent entries drawn from an arbitrary zero-mean distribution satisfying certain moment conditions.) GLMs include many important nonlinear estimation problems such as phase retrieval and logistic regression. Theorem 5.2 shows that the AMP algorithm and its performance characterization originally developed for GLMs with spatially coupled Gaussian designs [35] remain valid for a much broader class of designs.
To our knowledge, our scheme is the first one based on a random design to efficiently achieve recovery for QGT and pooled data in the linear regime with a sublinear number of tests. Moreover, the scheme is robust to noise. Numerical simulations show that the spatially coupled scheme outperforms the i.i.d. Bernoulli test design with AMP, as well as recovery algorithms based on convex programming.
Key technical ideas. Although the QGT model (1) is an instance of a linear model, an important constraint is that the test design matrix X can only contain binary entries. Therefore, we cannot apply the analysis from [34], which assumes a spatially coupled Gaussian design matrix. To prove the theoretical guarantees for our scheme, we reduce the SC-AMP algorithm to an abstract AMP iteration defined for any generalized white noise matrix. A state evolution result for this abstract AMP iteration was established by Wang et al. in [36], using which we obtain a rigorous asymptotic characterization of the SC-AMP algorithm (Theorem 3.2). To establish conditions for almost-exact recovery, we then need to analyze the fixed points of the SC-AMP state evolution. We do this in Theorem 3.5 via the potential function method [37], a powerful tool for characterizing the fixed points of coupled recursions. This characterization then yields the noise robustness and exactrecovery results (Theorem 3.7 and Corollary 3.8).
## 1.4 Other Related Work
Spatial coupling. Spatial coupling was introduced in coding theory as a means to construct LDPC codes that achieve capacity with an efficient belief propagation decoder [38, 39]. Spatial
coupling has since been applied in many estimation problems to improve on the performance of 'regular' (or i.i.d.) designs. For Boolean group testing in the sublinear regime (the number of defectives is p θ for θ ∈ (0 , 1) ), spatially coupled test designs enable efficient recovery with the asymptotically optimal number of tests, in both the noiseless [40] and noisy settings [41]. For QGT, Mashauri et al. [42, 43] investigated efficient schemes based on spatial coupled LDPC codes, and showed that they outperform previous constructions based on generalized LDPC codes [44,45].
Sublinear category regime. A few recent works have studied pooled data in the sublinear category regime, where one category is dominant with p -o p ( ) items, and the remaining ( L -1) categories have d = o p ( ) items. (In contrast, we consider the linear category regime, where the proportion of items in each category is Θ(1) , i.e., π l = Θ(1) for l ∈ [ L ] .) For the sublinear category regime, the information-theoretic lower bound for exact recovery is n = Ω( ) d tests [46, 47]. An efficient algorithm proposed in [46] achieves the lower bound when d = Θ( p κ ) , for a constant κ ∈ (0 , 1) . A lower complexity algorithm for the special case of QGT with d = Θ( p κ ) was recently proposed in [48]. For QGT in the sublinear regime, a number of recent works have proposed algorithms based on ideas from coding theory [42] and thresholding [49], which require Ω( d log p ) tests for exact recovery. Noisy versions of QGT were recently studied in [50] and [51], and QGT in the adaptive setting has been studied in [6,52].
## 2 Preliminaries
Notation. We write [ n ] := { 1 , . . . , n } and [ n : m ] for { n, n + 1 , . . . , m } where n < m . All vectors (including those corresponding to rows of matrices) are assumed to be column vectors unless otherwise stated. For a, b ∈ R n , ⟨ a, b ⟩ = a ⊤ b ∈ R is the inner product, a ⊙ b ∈ R n is the entry-wise product, and ⟨ a ⟩ = 1 n ∑ n i =1 a i denotes the empirical average of the entries of a . Matrices are denoted by upper case letters, and given a matrix A , we write A i, : for its i th row and A : ,j for its j th column. The operator norm is denoted by ∥ A ∥ op . For r ∈ [1 , ∞ ) and a vector a = ( a , . . . , a 1 n ) ∈ R n , we write ∥ a ∥ r for the ℓ r -norm, so that ∥ a ∥ r = ( ∑ n i =1 | a i | r ) 1 /r . We use 1 p to denote the vector of p ones, 0 p for the vector of p zeros, and I p for the p × p identity matrix. Given random variables U, V , we write U d = V to denote equality in distribution. We write ∂ f i ( ) · to denote the partial derivative of f with respect to (w.r.t.) the i th argument. Throughout, the function log( ) · has base e , and we use Bachmann-Landau asymptotic notation (i.e., O o , , Ω , ω , Θ ).
Almost-sure and Wasserstein convergence. Let { A n } be a sequence of random elements taking values in a Euclidean space E . We say that A n converges almost surely to a deterministic limit a ∈ E , and write A n a.s. → a , if P [lim n →∞ A n = a ] = 1 .
For a vector a ∈ R n and a random variable A ∈ R , we write a W r → A as n → ∞ , for the Wassersteinr convergence of the empirical distribution of the entries of a to the law of A . More generally, for vectors a , . . . , a 1 k ∈ R n and a random vector ( A , . . . , A 1 k ) ∈ R k , we write
$$a ^ { 1 }, \dots, a ^ { k } \stackrel { W _ { \bar { \gamma } } } { \rightarrow } ( A ^ { 1 }, \dots, A ^ { k } ) \text{ as } n \rightarrow \infty,$$
for the Wassersteinr convergence of the empirical distribution of rows of ( a , . . . , a 1 k ) ∈ R n × k to the joint law of ( A , . . . , A 1 k ) . This means that, for any continuous function ϕ : R k → R and input
Figure 1: The entries of X sc are independent with X sc ij ∼ Bernoulli ( αW r ( ) i , c ( j ) ) . Here W is an ( ω, Λ) base matrix with ω = 3 and Λ = 7 (see Definition 3.1). The white parts of X sc and W correspond to zeros.
vector ( a , . . . , a 1 i k i ) ∈ R k satisfying the polynomial growth condition [36]
$$| \phi ( a _ { i } ^ { 1 }, \dots, a _ { i } ^ { k } ) | \leq C ( 1 + \| ( a _ { i } ^ { 1 }, \dots, a _ { i } ^ { k } ) \| _ { 2 } ^ { r } ),$$
for a constant C > 0 , we have
$$\frac { 1 } { n } \sum _ { i = 1 } ^ { n } \phi ( a _ { i } ^ { 1 }, \dots, a _ { i } ^ { k } ) \to \mathbb { E } \left [ \phi ( A ^ { 1 }, \dots, A ^ { k } ) \right ] \text{ as } n \to \infty.$$
We write
$$a \stackrel { W } { \to } A ; \ \ ( a _ { 1 }, \dots, a _ { k } ) \stackrel { W } { \to } ( A _ { 1 }, \dots, A _ { k } ) \text{ as } n \to \infty,$$
to mean that the above Wassersteinr convergences hold for every order r ≥ 1 .
Model assumptions for QGT. The signal β ∈ { 0 1 , } p is independent of the design matrix. As n, p →∞ , we have n/p → δ > 0 , and the empirical distribution of the entries of the signal converges in Wasserstein distance to well-defined limits. More precisely, β W → ¯ β where ¯ β ∼ Bernoulli ( π ) . We note that the entries of β are not assumed to be independent or identically distributed.
## 3 Spatially Coupled Design for Quantitative Group Testing
The spatially coupled (SC) design matrix consists of independent Bernoulli entries whose parameters are specified by a base matrix W of dimension R × C . The SC design matrix is obtained by replacing each entry of the base matrix W rc by an n R × p C matrix with entries drawn independently from Bernoulli ( αW rc ) , where α, αW rc ∈ (0 , 1) . An example of a SC design matrix is shown in Figure 1. In this paper, we will use the following base matrix.
Definition 3.1. An ( ω, Λ) base matrix W is described by two parameters: the coupling width ω ≥ 1 and the coupling length Λ ≥ 2 ω -1 . The matrix has R = Λ+ ω -1 rows and C = Λ columns, with each entry indexed by ( r , c ) , for r ∈ [ R ] and c ∈ [ C ] . For α ≤ 0 5 . , the entries are given by
$$W _ { r c } = \begin{cases} \frac { 1 } { 2 \alpha } \left ( 1 - \sqrt { 1 - \frac { 4 \alpha ( 1 - \alpha ) } { \omega } } \right ) & \text{if $c\leq r\leq c+\omega-1$} \\ 0 & \text{otherwise.} \end{cases}$$
$$\text{For } \alpha > 0. 5, \text{ the non-zero entries } ( \text{when } c \leq r \leq c + \omega - 1 ) \text{ are given by } \frac { 1 } { 2 \alpha } \left ( 1 + \sqrt { 1 - \frac { 4 \alpha ( 1 - \alpha ) } { \omega } } \right ).$$
Figure 1 shows an ( ω, Λ) base matrix with ω = 3 and Λ = 7 . The spatially coupled (SC) design matrix, denoted by X sc , has entries
$$X _ { i j } ^ { s c } \underset { \sim } { \text{emodii} } \text{ Bernoulli} ( \alpha W _ { r ( i ) \epsilon ( j ) } ), \ \ i \in [ n ], \, j \in [ p ].$$
for some fixed constant α ∈ (0 , 1) . Here the operators r ( ) · : [ n ] → [ R ] and c ( ) · : [ p ] → [ C ] map a particular row or column index to its corresponding row block or column block index in W . The band-diagonal structure of ( ω, Λ) base matrix here is similar to the ones used for SC sparse regression codes [53] and for SC generalized linear models [35], but the values of the non-zero entries are different. Here the base matrix specifies the Bernoulli parameters for each block of the design, whereas in [35,53] it specifies the variances for the Gaussian entries in each block.
With the i.i.d. design, the number of defectives per test has expected value απp and standard deviation √ απ (1 -απ p ) . Similarly, for the spatially coupled design, it can be verified that the number of defective items per test has mean that is linear in p and standard deviation of order √ p . Since the fluctuation around the mean contains the useful information in each test, and because AMP requires a design matrix with zero-mean entries, we recenter and rescale the data before applying the AMP algorithm. We now describe this preprocessing of the data, which was also done in [14] for i.i.d. designs.
The i.i.d. design matrix, denoted by X iid , has each entry sampled i.i.d. ∼ Bernoulli ( α ) , for some fixed constant α ∈ (0 , 1) . Note that the i.i.d. matrix is a special case of the SC matrix, with R = C = 1 and W = 1 . A key difference between the i.i.d. design and the SC design (with an ( ω, Λ) base matrix) is that the latter includes many fewer items in each test. Indeed, with the i.i.d. design each item is included in a test with probability α , whereas in the SC design, each test includes items from at most ω adjacent column blocks, each with p/ C items (see Figure 1). In the SC design, since the tests corresponding to the first and last row blocks involve the fewest items, the corresponding entries of β are easier to recover than the others. A good estimate for these entries helps the algorithm recover the entries in the adjacent blocks, creating a decoding wave that propagates from the ends towards the center.
The rescaled i.i.d. matrix, denoted by ˜ X iid , is defined as
$$\widetilde { X } ^ { \text{iid} } = \frac { X ^ { \text{iid} } - \alpha 1 _ { n } 1 _ { p } ^ { \top } } { \sqrt { n \alpha ( 1 - \alpha ) } }.$$
We note that ˜ X iid has independent entries with E [ ˜ X ij ] = 0 and Var[ ˜ X ij ] = 1 /n .
The rescaled spatially coupled (SC) matrix ˜ X sc is defined as follows. For i ∈ [ n , j ] ∈ [ p ] , using the shorthand r ≡ r ( ) i , c ≡ c ( j ) , its entries are given by
$$\widetilde { X } _ { i j } ^ { s c } = \frac { X _ { i j } ^ { s c } - \alpha W _ { r c } } { \sqrt { n \alpha ( 1 - \alpha ) / R } } = \begin{cases} \frac { 1 - \alpha W _ { r c } } { \sqrt { n \alpha ( 1 - \alpha ) / R } } & \text{with probability $\alpha W_{rc}$,} \\ \frac { - \alpha W _ { r c } } { \sqrt { n \alpha ( 1 - \alpha ) / R } } & \text{with probability $1-\alpha W_{rc}$.} \end{cases}$$
It is straightforward to verify that E [ ˜ X sc ij ] = 0 , Var[ ˜ X sc ij ] = R W rc (1 -αW rc ) n (1 -α ) . In particular, for the ( ω, Λ) base matrix in Definition 3.1, we have
$$\text{Var} [ \widetilde { X } _ { i j } ^ { s c } ] = \begin{cases} \frac { R } { n \omega } & \text{if $c\leq r\leq c+\omega-1$} \\ 0 & \text{otherwise.} \end{cases}$$
Rewriting QGT. The AMP algorithm and its analysis require the design matrix to have independent zero-mean entries, so we recenter and rescale the QGT model in (1) to express it in terms of the rescaled design ( ˜ X sc or ˜ X iid ). For the SC design, using (10), we have for i ∈ [ n ] :
$$\overset { \cdot } { \text{of the rescaled design } ( \widetilde { X } ^ { s } \text{$c$} \text{ or $\widetilde{X}^{iid}$} ). \text{ For the SC design, using } (10), we have for $i\in [n] \colon \\ y _ { i } = \overset { p } { \underset { j = 1 } { \underset { k = 1 } { \sum } } } X ^ { s c } _ { i j } \beta _ { j } + \Psi _ { i } = \overset { p } { \underset { j = 1 } { \sum } } \left ( \alpha W _ { r ( i ) c ( j ) } + \sqrt { \frac { n \alpha ( 1 - \alpha ) } { R } } \widetilde { X } ^ { s c } _ { i j } \right ) \beta _ { j } + \Psi _ { i } \\ \Longrightarrow & y _ { i } - \overset { p } { \underset { j = 1 } { \sum } } \alpha W _ { r ( i ) c ( j ) } \beta _ { j } = \sqrt { \frac { n \alpha ( 1 - \alpha ) } { R } } \overset { p } { \underset { j = 1 } { \sum } } \widetilde { X } ^ { s c } _ { i j } \beta _ { j } + \Psi _ { i } \\ \Longrightarrow & \frac { 1 } { \sqrt { n \alpha ( 1 - \alpha ) / R } } \left ( y _ { i } - \alpha \left ( W _ { r ( i ) 1 } \underset { j \in \mathcal { J } _ { 1 } } { \underset { k = 1 } { \sum } } \beta _ { j } + \cdots + W _ { r ( i ) c } \sum _ { j \in \mathcal { J } _ { c } } \beta _ { j } \right ) \right ) = ( \widetilde { X } ^ { s c } _ { i, i } ) ^ { \top } \beta + \frac { \Psi _ { i } } { \sqrt { n \alpha ( 1 - \alpha ) / R } }, \\ \text{where $\mathcal{J}_{c}=[(c-1)p/C+1\coloncp/C]$ for $c\in[C]$. Denoting the left-hand side above by}$$
where J c = ( [ c -1) p/ C +1 : c p/ C ] for c ∈ [ C ] . Denoting the left-hand side above by
$$\dot { \vec { y } } _ { i } \coloneqq \frac { 1 } { \sqrt { n \alpha ( 1 - \alpha ) / R } } \Big ( y _ { i } - \alpha \left ( W _ { r ( i ) 1 } \sum _ { j \in \mathcal { J } _ { 1 } } \beta _ { j } + \cdots + W _ { r ( i ) C } \sum _ { j \in \mathcal { J } _ { c } } \beta _ { j } \right ) \right ),$$
gives us the rescaled QGT model:
$$\tilde { y } _ { i } = ( \widetilde { X } _ { i, ; } ^ { \text{sc} } ) ^ { \text{T} } \beta + \widetilde { \Psi } _ { i }, \quad \text{with} \quad \widetilde { \Psi } _ { i } \coloneqq \frac { \Psi _ { i } } { \sqrt { n \alpha ( 1 - \alpha ) / R } } \,, \quad i \in [ n ].$$
The term ∑ j ∈J c β j in (12) is the number of defective items in the sub-vector of β indexed by J c , for c ∈ [ C ] . In the noiseless setting, the terms ∑ j ∈J c β j can be obtained with an extra C tests, where we only include items from J c in the c th test, for c ∈ [ C ] . The extra C = O (1) tests does not affect our results since the limiting sampling ratio lim n →∞ n/p = δ remains the same. Since 1 p ∑ j ∈J c β j a.s. → π C for p →∞ via the strong law of large numbers, so we can also estimate ∑ j ∈J c β j using p C π . (However the error in this estimate would be of order √ p/ C .)
For an i.i.d. design X iid , we similarly recenter and rescale the QGT model to express it in terms of the rescaled i.i.d. matrix ˜ X iid in (9). For i ∈ [ n ] , we have
$$y _ { i } = \left ( X _ { i, i } ^ { i \text{id} } \right ) ^ { \top } \beta + \Psi _ { i } \implies & \tilde { y } _ { i } = \left ( \widetilde { X } _ { i, i } ^ { \text{id} } \right ) ^ { \top } \beta + \widetilde { \Psi } _ { i }, \\ & \quad \text{where } \tilde { y } _ { i } \coloneqq \frac { y _ { i } - \alpha d } { \sqrt { n \alpha ( 1 - \alpha ) } }, \quad \widetilde { \Psi } _ { i } = \frac { \Psi _ { i } } { \sqrt { n \alpha ( 1 - \alpha ) } }, \quad \quad ( 1 4 )$$
where d is the number of defective items.
$$\dots$$
Choice of α . For a spatially coupled design X sc constructed from an ( ω, Λ) base matrix, recall from (10) that the rescaled matrix ˜ X sc has independent zero-mean Bernoulli entries with variances given by (11). Notice that the distribution of ˜ X sc does not depend on α . Hence, in the rescaled model, α only affects the variance of the noise ˜ Ψ , which is minimized when α = 0 5 . . We therefore use α = 0 5 . for all our experiments.
Taking α = 0 5 . , it is useful to compare the SC design with the i.i.d. Bernoulli( 0 5 . ) design. (For the i.i.d. design, taking the Bernoulli parameter to be 0 5 . is optimal with respect to both the information-theoretic limits [11] and efficient recovery via AMP [14].) Taking α = 0 5 . in Definition 3.1, we have that the non-zero blocks of the SC design are drawn independently from an Bernoulli distribution with parameter 1 2 ( 1 -√ 1 -1 /ω ) . From structure of X sc (see Figure 1), it follows that the expected number of items included in each test is at most p ω C 2 ( 1 -√ 1 -1 /ω ) . For large ω , this is approximately p 4 C items per test. In contrast, for the Bernoulli( 0 5 . ) design, the expected number of items per test is p 2 . Thus, for large C , the expected number of items per test is much smaller for the SC design than the i.i.d. design.
Noise scaling assumption. In the rescaled QGT models (13) and (14), it can be verified that the terms ( ˜ X sc i, : ) ⊤ β and ( ˜ X iid i, : ) ⊤ β have zero mean and variance of constant order, for each i ∈ [ n ] . Therefore, for the rescaled model to be meaningful, the noise ˜ Ψ i should also have a mean and variance of constant order. This is guaranteed by the following assumption. The empirical distribution of rescaled noise vector ˜ Ψ in (13) converges to a well defined limit. More precisely, there exists, ¯ Ψ ∼ P ¯ Ψ with E [ ¯ Ψ 2 ] =: σ 2 < ∞ , such that ˜ Ψ W → ¯ Ψ as n →∞ . We emphasize that the base matrix parameter R is fixed as n →∞ . A similar distributional assumption holds for the rescaled noise vector with the i.i.d. design in (14).
## 3.1 SC-AMP Algorithm
Consider the rescaled SC model (13). Given ( ˜ X ,y sc ˜) , the SC-AMP algorithm iteratively produces estimates ˆ β k for k ≥ 1 . For iteration k ≥ 0 , the algorithm computes:
$$\widetilde { \Theta } ^ { k } & = \tilde { y } \, - \, \widetilde { X } ^ { s c } \hat { \beta } ^ { k } \, + \, b ^ { k } \odot Q ^ { k - 1 } \odot \widetilde { \Theta } ^ { k - 1 } \, \in \, \mathbb { R } ^ { n } \\ \beta ^ { k + 1 } & = ( \widetilde { X } ^ { s c } ) ^ { \top } ( Q ^ { k } \odot \widetilde { \Theta } ^ { k } ) \, - \, c ^ { k } \odot \hat { \beta } ^ { k } \, \in \, \mathbb { R } ^ { p }, \quad \hat { \beta } ^ { k + 1 } = f _ { k + 1 } ( \beta ^ { k + 1 }, \mathcal { C } ) \, \in \, \mathbb { R } ^ { p },$$
where f k : R 2 → R acts row-wise on the inputs ( β , k C ) and
$$\mathcal { C } \coloneqq ( \underbrace { 1, \dots, 1 } _ { p / C \text{ entries } p / C \text{ entries } } \,, \, \underbrace { 2, \dots, 2 } _ { p / C \text{ entries } } \,, \, \dots, \, \underbrace { C, \dots, C } _ { p / C \text{ entries } } ) ^ { \top } \in \mathbb { R } ^ { p }.$$
The algorithm is initialized with ˆ β 0 = E [ ¯ ]1 β p and ˜ Θ = ˜ 0 y -˜ X β sc ˆ 0 . We recall that ⊙ is the Hadamard (entry-wise) product. To define b k , we use the partitions [ p ] = ⋃ C c =1 J c and [ n ] = ⋃ R r =1 I r where
$$\mathcal { J } _ { c } = \left \{ ( c - 1 ) \frac { p } { C } + 1, \dots, c \frac { p } { C } \right \} \text{ for } c \in [ C ], \quad \mathcal { I } _ { \tau } = \left \{ ( r - 1 ) \frac { n } { R } + 1, \dots, r \frac { n } { R } \right \} \text{ for } r \in [ R ]. \quad ( 1 7 )$$
Then, letting
$$\widetilde { W } _ { \text{rc} } \coloneqq \frac { W _ { \text{rc} } ( 1 - \alpha W _ { \text{rc} } ) } { 1 - \alpha } = \begin{cases} 1 / \omega & \text{if $c\leq r\leq c+\omega-1$} \\ 0 & \text{otherwise.} \end{cases},$$
the entries of b k ∈ R n are
$$b _ { i } ^ { k } = \sum _ { c = 1 } ^ { c } \frac { \widetilde { W } _ { r c } } { n / R } \sum _ { j \in \mathcal { J } _ { c } } \partial _ { 1 } f _ { k } ( \beta _ { j } ^ { k }, \mathbf c ), \ \ i \in [ n ].$$
The second equality in (18) follows from Definition 3.1. The function f k and the vectors Q ,c k k in (15) will be defined later in (19) and (23), respectively, in terms of state evolution parameters. We note that the time complexity of each iteration in (15) is O np ( ) .
We introduce some additional notation to handle the block-wise structure of the SC-AMP iterates. For c ∈ [ C ] and r ∈ [ R ] , define β c := β J c ∈ R p/ C and ˆ β k c := ˆ β k J c ∈ R p/ C to be the c th blocks of β ∈ R p and ˆ β k ∈ R p respectively, and Θ := Θ r I r ∈ R n/ R and ˜ Θ k r := Θ ˜ k I r ∈ R n/ R to be the r th blocks of Θ := ˜ X β sc ∈ R n and ˜ Θ k ∈ R n respectively. Similar notation simplifications will be used for other vectors where c and r will replace J c and I r in the subscripts of the vectors.
State evolution. The 'memory' terms b k ⊙ Q k -1 ⊙ ˜ Θ k -1 and -c k ⊙ ˆ β k in (15) debias the iterates ˜ Θ k and β k +1 , ensuring that their empirical distributions are accurately captured by state evolution in the high-dimensional limit. These iterates have a block-wise distributional structure. Recall from the model assumptions that the empirical distribution of the signal β converges to the law of ¯ β ∼ Bernoulli ( π ) . Theorem 3.2 below shows that for each k ≥ 1 and c ∈ [ C ] , the empirical distribution of β k c converges to the distribution of ( χ k c ) 2 ¯ + β χ G k c , where G ∼ N (0 1) , is a standard Gaussian independent of ¯ β , and the deterministic parameter χ k c ∈ R is defined below via the state evolution recursion. Thus, f k in (15) can be viewed as a denoising function that can be tailored to take advantage of the prior on ¯ β . Indeed, we choose f k in this manner, and for k ≥ 1 , define:
$$f _ { k } ( s, c ) = \mathbb { E } [ \bar { \beta } \, | \, ( \chi _ { \epsilon } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { \epsilon } ^ { k } G = s ] \,, \quad \text{for } c \in [ C ].$$
State evolution iteratively computes the parameter χ k c ∈ R as follows, for k ≥ 1 . Letting
$$\delta _ { \text{in} } = \frac { n / R } { p / C } = \frac { C } { R } \delta \,,$$
given χ k c for c ∈ [ C ] , we compute χ k +1 c as:
$$( \chi _ { c } ^ { k + 1 } ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \left ( \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \sum _ { c ^ { \prime } = 1 } ^ { C } \widetilde { W } _ { r c ^ { \prime } } \mathbb { E } \left [ \left ( \bar { \beta } - f _ { k } \left ( \chi _ { c ^ { \prime } } ^ { k } \right ) ^ { 2 } \bar { \beta } + \chi _ { c ^ { \prime } } ^ { k } G, \, c ^ { \prime } \right ) \right ) ^ { 2 } \right ] \right ) ^ { - 1 },$$
where G ∼ N (0 1) , is independent of ¯ β , and σ 2 is the variance defined in the noise scaling assumption on p.9. The recursion is initialized with
$$( \chi ^ { 1 } _ { c } ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \left ( \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \sum _ { c ^ { \prime } = 1 } ^ { C } \widetilde { W } _ { r c ^ { \prime } } \text{Var} ( \bar { \beta } ) \right ) ^ { - 1 }, \ \ c \in [ C ].$$
The vectors Q k ∈ R n and c k ∈ R p in (15) have a block-wise structure, and are defined as follows. For i ∈ [ n ] , j ∈ [ p ] , recalling that r ( ) i and c ( j ) denote the respective row-block and column-block indices, we have:
$$Q _ { i } ^ { k } = \left ( \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { c } \widetilde { W } _ { r ( i ) c } \mathbb { E } \left [ \left ( \bar { \beta } - f _ { k } ( \left ( \chi _ { c } ^ { k } \right ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G, \, c \right ) ) ^ { 2 } \right ] \right ) ^ { - 1 }, \quad c _ { j } ^ { k } = - ( \chi _ { c ( j ) } ^ { k + 1 } ) ^ { 2 }.$$
The SC-AMP algorithm in in (15) and its state evolution are equivalent to the one proposed for a spatially coupled Gaussian design [34]. The key difference is that our algorithm uses a rescaled spatially coupled Bernoulli design. The theorem below shows that the state evolution guarantees remain valid for this setting. For an i.i.d. design (where R = C = 1 ), SC-AMP reduces to the standard AMP algorithm [16] for an i.i.d. Gaussian design.
Theorem 3.2 (State evolution result for SC-AMP) . Consider the QGT model (1) with a spatially coupled design defined via the ( ω, Λ) base matrix in Definition 3.1. Let the model assumptions in Section 2 and the noise scaling assumption (p. 9) be satisfied. Then, for the SC-AMP algorithm in (15) , run on the recentered and rescaled QGT model (13) with the denoising functions f k in (19) , we have the following convergence guarantee. For each k ≥ 0 and c ∈ [ C ] :
$$( \beta _ { \mathrm c }, \beta _ { \mathrm c } ^ { k + 1 } ) \stackrel { W _ { 2 } } { \to } ( \bar { \beta }, ( \chi _ { \mathrm c } ^ { k + 1 } ) ^ { 2 } \bar { \beta } + \chi _ { \mathrm c } ^ { k + 1 } G )$$
almost surely as n, p →∞ with n/p → δ .
The theorem is proved in Section 5, where the SC-AMP algorithm is shown to be a special case of an AMP algorithm for a generalized linear model with a generic spatially coupled design. We prove Theorem 3.2 by establishing a state evolution result for this general AMP algorithm (Theorem 5.2).
Performance measures. Theorem 3.2 allows us to compute the limiting values of performance measures such as the mean-squared error (MSE) and the normalized squared correlation, via the convergence property in (6). The MSE of the AMP estimates ˆ β k satisfy the following almost surely, for k ≥ 1 :
$$\lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \mathbb { E } \left [ \left ( \bar { \beta } - f _ { k } ( ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G, c ) \right ) ^ { 2 } \right ].$$
And the normalized squared correlation of the AMP estimates ˆ β k satisfy:
$$\lim _ { p \to \infty } \frac { \langle \hat { \beta } ^ { k }, \beta \rangle } { \| \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } \cdot \| \beta \| _ { 2 } ^ { 2 } } = \frac { ( \frac { 1 } { \bar { c } } \sum _ { c = 1 } ^ { C } \mathbb { E } [ f _ { k } ( ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G, c ) \cdot \bar { \beta } ] ) ^ { 2 } } { ( \frac { 1 } { \bar { c } } \sum _ { c = 1 } ^ { C } \mathbb { E } [ f _ { k } ( ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G, c ) ^ { 2 } ] ) \cdot ( \mathbb { E } [ \bar { \beta } ^ { 2 } ] ) }, \quad k \geq 1.$$
We can also obtain formulas for the limiting values of the false positive rate (FPR) and false negative rate (FNR). The choice of f k in (19) outputs a vector in R p , but we can obtain an estimate in { 0 1 , } p by thresholding the AMP iterate ˆ β K in the final iteration K to output a hard decision. For some chosen constant ζ , let us define the hard decision to be
$$1 \left \{ \hat { \beta } _ { j } ^ { K } > \zeta \right \} = 1 \left \{ f _ { K } ( \beta _ { j } ^ { K }, \mathfrak { c } ) > \zeta \right \}, \quad \text{for } j \in \mathcal { J } _ { \mathfrak { c } }, \mathfrak { c } \in [ \mathbb { C } ].$$
where the indicator function is applied component-wise to ˆ β k c . That is, we declare large entries of ˆ β K to be one (i.e., defective) and small entries of ˆ β K to be zero (i.e., non-defective). Based on the above function, let us denote the estimated defective set as ̂ S = { j : ˆ β K j > ζ } .
The false positive rate (FPR) and the false negative rate (FNR) are defined as:
$$\ F P R = \frac { \sum _ { j = 1 } ^ { p } 1 \{ \beta _ { j } = 0 \cap j \in \widehat { \mathcal { S } } \} } { p - \sum _ { j = 1 } ^ { p } \beta _ { j } } \quad \text{and} \quad \ F N R = \frac { \sum _ { j = 1 } ^ { p } 1 \{ \beta _ { j } = 1 \cap j \notin \widehat { \mathcal { S } } \} } { \sum _ { j = 1 } ^ { p } \beta _ { j } }.$$
Corollary 3.3. Under the same assumptions as for Theorem 3.2, with a threshold ζ ∈ [0 , 1] for the final iteration K , as p →∞ , we have
$$\ F P R \stackrel { a. s. } { \rightarrow } \frac { 1 } { \overline { c } } \sum _ { c = 1 } ^ { c } \mathbb { P } [ f _ { K } ( \chi _ { \epsilon } ^ { K } G, c ) > \zeta ] \ \ a n d \ F N R \stackrel { a. s. } { \rightarrow } \frac { 1 } { \overline { c } } \sum _ { c = 1 } ^ { c } \mathbb { P } [ f _ { K } ( ( \chi _ { \epsilon } ^ { K } ) ^ { 2 } + \chi _ { \epsilon } ^ { K } G, c ) \leq \zeta ].$$
The result follows from Theorem 3.2 by applying the convergence property to suitable indicator functions. The proof uses the same steps as the analogous result for i.i.d. Bernoulli designs [14, Corollary 5.2], and is omitted.
## 3.2 Almost-Exact Recovery
Given ˆ β k , the SC-AMP estimate after k iterations, let us define the quantized estimate to be
$$\tilde { \beta } _ { j } ^ { k } = \begin{cases} 1 & \text{if $\hat{\beta}_{j}^{k}>0.5$} \\ 0 & \text{otherwise,} \end{cases}$$
̸
In this section, we show that the SC-AMP algorithm can attain almost-exact recovery with n = o p ( ) tests, by proving that it attains almost-exact recovery for any δ > 0 (recall that δ = lim n →∞ n p ). To this end, we introduce the potential function to analyze the asymptotic MSE of SC-AMP as k → ∞ . Potential functions are widely used to characterize the limiting MMSE and mutual information in high-dimensional estimation problems (see, e.g., [54, 55]). Here we will use it only to characterize the asymptotic MSE of the AMP algorithm, both with and without spatial coupling (see Theorem 3.5).
Then, recalling the almost-exact recovery criterion in (2), the SC-AMP algorithm achieves almostexact recovery if lim k →∞ lim p →∞ 1 p ∑ p j =1 1 { ˜ β k j = β j } = 0 .
Definition 3.4. For b ∈ [0 , Var( ¯ )] β , δ > 0 , the scalar potential function for the rescaled QGT model is defined as
$$U ( b ; \delta ) \coloneqq - \delta \Big ( 1 - \frac { \sigma ^ { 2 } } { ( b / \delta ) + \sigma ^ { 2 } } \Big ) + \delta \log \Big ( 1 + \frac { b } { \delta \sigma ^ { 2 } } \Big ) + 2 I \Big ( \bar { \beta } ; \sqrt { \frac { 1 } { ( b / \delta ) + \sigma ^ { 2 } } } \bar { \beta } + G \Big ). \quad \ ( 3 1 )$$
Here the mutual information I ( ; · · ) is computed with ¯ β ∼ P ¯ β independent of G ∼ N (0 1) , , and σ 2 = E [ ¯ Ψ 2 ] is the second moment of the rescaled noise (see p. 9).
Figure 2 plots the potential function for various values of δ . The next theorem characterizes the limiting MSE of the AMP algorithm via the minimizers and stationary points of the potential function. For clarity, we refer to the AMP algorithm under the rescaled i.i.d. QGT model in (14) as the iid-AMP algorithm, and the AMP algorithm under the rescaled spatially coupled QGT model in (13) as the SC-AMP algorithm. The SC-AMP algorithm in (15) reduces to iid-AMP with the trivial base matrix ( R = C = 1 ).
Theorem 3.5 (MSE of SC-AMP and iid-AMP) . Consider the QGT model (1) , and let the model assumptions in Section 2 and the noise scaling assumption (p. 9) be satisfied.
Figure 2: U b δ ( ; ) vs. b , for different δ . π = 0 1 . and σ = 1 × 10 -30 .
- 1. Consider a spatially coupled design defined via an ( ω, Λ) base matrix (Definition 3.1). For any γ > 0 , there exist ω 0 < ∞ and k 0 < ∞ such that for all ω > ω 0 and k > k 0 , the asymptotic MSE of the SC-AMP algorithm almost surely satisfies:
$$\lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } < \left ( \max \left \{ \underset { b \in [ 0, V \arrow ( \bar { \beta } ) ] } { \arg \min } U ( b ; \delta _ { \text{in} } ) \right \} + \gamma \right ) \frac { \Lambda + \omega } { \Lambda }.$$
- 2. With an i.i.d. design (i.e., 1 × 1 base matrix with W 11 = 1 ), the asymptotic MSE of the iid-AMP algorithm almost surely satisfies:
$$\lim _ { k \to \infty } \lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } = \max \left \{ b \in [ 0, \text{Var} ( \bar { \beta } ) ] \, \colon \partial _ { 1 } U ( b ; \delta ) = 0 \right \},$$
where ∂ 1 denotes the partial derivative w.r.t. the first argument.
The proof is given in Section 6. Part 1 of the theorem says that for sufficiently large base matrix parameters (with ω ≪ Λ ), the MSE of the SC-AMP algorithm is bounded by the largest minimizer of the potential function. (The max {·} indicates that if there are multiple minimizers, the largest one is chosen.) Part 2 of the theorem says that the MSE of iid-AMP algorithm is given by the largest stationary point of the potential function. In Figure 2, we observe that for δ = 0 05 . and δ = 0 02 . , the unique minimizer is b = 0 , but the largest stationary point is strictly larger than zero. This implies that the limiting MSE of SC-AMP algorithm is 0, but that of iid-AMP algorithm is strictly larger than 0. The next lemma quantifies this observation, showing that for any δ > 0 , the largest minimizer of the potential function tends to zero as the noise variance σ → 0 .
Lemma 3.6. Consider the scalar potential function U b δ ( ; ) in (31) with δ > 0 . For any ∆ ∈ (0 , δ ) , there exists σ 0 (∆) > 0 such that for all σ < σ 0 (∆) , we have the rate
$$\max \left \{ \underset { b \in ( 0, \text{Var} ( \bar { \beta } ) ] } { \arg \min } U ( b ; \delta ) \right \} < \frac { 7 } { 2 } \delta \left ( \sigma ^ { 2 - 2 \Delta / \delta } \right ).$$
The proof is given in Appendix B. Using Lemma 3.6 in Part 1 of Theorem 3.5 yields the following bound on the MSE of the SC-AMP algorithm in the low noise regime.
Theorem 3.7 (MSE of SC-AMP in the low-noise regime) . Consider the setup of part 1 of Theorem 3.5, for any δ > 0 . Then for any ϵ > 0 and ∆ ∈ (0 , δ/ (1 + ϵ )) , there exists σ 0 (∆) > 0 such that following holds for any noise variance σ < σ 0 (∆) . There exist finite ω 0 and k 0 such that for all ω > ω 0 , k > k 0 , and Λ sufficiently large, the asymptotic MSE of the SC-AMP algorithm almost surely satisfies:
$$\lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } < 4 \delta \sigma ^ { 2 - 2 \Delta ( 1 + \epsilon ) / \delta } + \epsilon.$$
Proof. Let ϵ, γ > 0 . Recalling that δ in = δ Λ Λ+ ω -1 , for sufficiently large Λ /ω we have δ in < δ 1+ ϵ . Then, from Part 1 of Theorem 3.5, for k > k 0 , ω > ω 0 and sufficiently large Λ /ω , we have
$$\lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } < \left ( \max \left \{ \underset { b \in [ 0, V \bar { \epsilon } ] } { \arg \min } U ( b ; \delta _ { \text{in} } ) \right \} + \gamma \right ) ( 1 + \epsilon ),$$
Taking γ to be small enough and using Lemma 3.6 in (35) gives the result.
Corollary 3.8 (SC-AMP achieves almost exact recovery for any δ > 0 ) . Consider the setting and assumptions of Theorem 3.7. Let ˜ β k be the quantized estimate produced from the SC-AMP iterate ˆ β k , according to (30) . Then, almost surely we have:
̸
$$\lim _ { p \to \infty } \frac { 1 } { p } \sum _ { j = 1 } ^ { p } \mathbf 1 \{ \tilde { \beta } _ { j } ^ { k } \neq \beta _ { j } \} \leq 4 \lim _ { p \to \infty } \frac { 1 } { p } \| \beta - \hat { \beta } ^ { k } \| _ { 2 } ^ { 2 } < 4 \left ( 4 \delta \sigma ^ { 2 - 2 \Delta ( 1 + \epsilon ) / \delta } + \epsilon \right ).$$
̸
In particular, for noiseless QGT we have lim k →∞ lim p →∞ 1 p ∑ p j =1 1 { ˜ β k j = β j } = 0 almost surely for any δ > 0 .
̸
̸
̸
Proof. From the definition of the quantized estimate in (30), it follows that for j ∈ [ p ] , we have | β j -ˆ β k j | ≥ (0 5) . 1 { ˜ β k j = β j } , which implies 1 { ˜ β k j = β j } ≤ 4 | β j -ˆ β k j | 2 . This gives the first inequality in (36). The second equality follows from Theorem 3.7. The result for the noiseless case follows by setting σ = 0 , and taking a sequence ( ϵ k ) such that ϵ k → 0 as k → ∞ . We note that lim p →∞ 1 p ∑ p j =1 1 { ˜ β k j = β j } exists for each k by the state evolution result in Theorem 3.2.
The guarantees in Corollary 3.8 are analogous to those in [34, Theorem 1.7 and Corollary 1.8] for a linear model with a spatially coupled Gaussian design matrix. Specifically, [34, Theorem 1.7] shows that when the sampling ratio δ is larger than the Rényi information dimension of the signal prior, the MSE of the SC-AMP algorithm satisfies lim k →∞ lim p →∞ 1 p ∥ β -ˆ β k ∥ 2 2 < Cσ 2 , for sufficiently small noise variance σ 2 . Here the constant C depends on δ and on the prior. The Rényi information dimension for a Bernoulli prior is 0, so Corollary 3.8 is consistent with the result in [34]. The key difference is that we use a binary-valued SC design for the QGT model rather than the Gaussian one in [34]. Our analysis of the fixed point of the SC state evolution to establish Theorem 3.5 is also simpler than in [34], where the authors use a continuum version of the state evolution along with a perturbation argument. In contrast, we use a straightforward potential function analysis based on the recipe provided in [37] for analyzing coupled recursions.
(a) SC-AMP vs. iid-AMP
Figure 3: Normalized squared correlation for Noiseless QGT. π = 0 3 . , p = 20000 . With spatial coupling parameters ω = 6 Λ = 40 , , the inner block size p/ Λ = 500 . In (b), we use p = 2000 for both iid-LP and SC-LP due to computational constraints. Error bars indicate one standard deviation.
## 3.3 Numerical Simulations
We present simulation results for finite length SC-AMP and compare its performance against alternative algorithms and the information-theoretic limit. The performance in all the plots is either measured via the normalized squared correlation between the SC-AMP estimate and the signal (see (26)) or via the FPR and FNR (see (28)). In the plots, curves labeled 'SC-AMP' show the empirical performance of the SC-AMP algorithm, while the 'SC-SE' curves refer to its theoretical performance predicted via state evolution. The corresponding curves for an i.i.d. design are labeled 'iid-AMP' and 'iid-SE'. For empirical performance curves, each point is obtained from 10 independent runs, where in each run, the SC-AMP algorithm is executed for 300 iterations. Other implementation details are described in Appendix C. Python code for all the simulations is available at [56].
Figure 3b shows how the AMP algorithm compares to linear programming (LP) estimator,
Figure 3a shows how normalized squared correlation varies with the sampling ratio δ for noiseless QGT, for both spatially coupled and i.i.d. designs. We observe that SC-AMP outperforms iid-AMP, justifying the use of the SC design. The orange vertical line show the information-theoretic lower bound on the ratio n/p obtained from (4). Specializing (4) to the case of L = 2 , we get the information-theoretic lower bound on the number of tests for noiseless QGT: n ∗ = 2 H π ( ) p log p . We observe that the performance of SC-SE improves and approaches n ∗ as we increase the size of the spatial coupling parameters ( ω, Λ) from (6,40) to (40,400). We did not implement the SC-AMP for ( ω, Λ) = (40 400) , as it requires a large amount of computational memory. The difference between the SC-SE plot and the SC-AMP plot for (6 , 40) is due to finite length effects, since the inner block size p/ Λ is only 500.
defined as the solution of the following linear program:
$$\minimize & \ \| \beta \| _ { 1 } \\ \text{subject to} & \ y = X \beta, \text{ and } 0 \leq \beta _ { j } \leq 1, \ \ j \in [ p ].$$
Similar reconstruction algorithms are commonly used for compressed sensing [57]. LP based estimators have also been used in Boolean in group testing [1]. We observe that the AMP algorithm outperforms LP for both i.i.d. and SC designs, and that the performance of LP is similar with both designs. This is because the LP algorithm is not equipped to take advantage of the spatially coupled design. LP is also more computationally intensive than the SC-AMP algorithm and challenging to implement for large values of p . Therefore, we use a smaller p for all our LP experiments.
Figure 4b shows the tradeoff between the FPR and the FNR for noisy QGT with δ = 0 46 . and σ 2 = 0 0016 . . Following the model in (1), for the i.i.d. design we consider Gaussian noise with Ψ i i . i . d . ∼ N (0 , pσ 2 ) , as previously investigated in [11, 14]. For the SC design, we consider Ψ i i . i . d . ∼ N (0 , pσ 2 / (2 C ) ) . As described in Section 3, for α = 0 5 . , the expected number of items in each test is approximately p 4 C for the SC design, compared to p/ 2 for the i.i.d. design. This choice of noise variance for the SC model ensures that the signal-to-noise ratio E [ ∥ Xβ ∥ 2 ] / E [ ∥ Ψ ∥ 2 ] is similar for both designs.
Figure 4a shows the tradeoff between the FPR and the FNR for noiseless QGT with δ = 0 38 . . The tradeoff curve is obtained by thresholding the AMP or LP estimate with different thresholds ζ , as described in (27). SC-AMP achieves perfect recovery at this value of δ , so its FPR and FNR are both 0, for all threshold values. As expected, SC-AMP does significantly better than iid-AMP and LP.
In the noisy setting, the AMP algorithm is compared to the following convex programming (CVX) estimator:
$$\minimize & \quad \frac { 1 } { 2 \text{Var} ( \Psi ) } \| y - X \beta \| _ { 2 } ^ { 2 } + \| \beta \| _ { 1 } \log \frac { 1 - \pi } { \pi } \\ \text{subject to} & \quad 0 \leq \beta _ { j } \leq 1, \quad j \in [ p ].$$
This estimator is obtained via a convex relaxation of the MAP estimator for QGT. Figure 4b shows that in the presence of a small amount of noise, SC-AMP continues to achieve perfect recovery, outperforming both CVX and iid-AMP. Surprisingly, in the presence of noise, the performance of CVX is worse with the SC design than with the i.i.d. one, possibly because it does not take advantage of the band-diagonal structure in the SC design matrix.
## 4 SC-AMP for Pooled Data
In this section we extend the SC-AMP algorithm to the pooled data model in (3) with the spatially coupled design X sc defined in (8). We apply SC-AMP to a centered and rescaled version of the pooled model, as we did for QGT in (13). Recalling the decomposition X sc ij = αW rc + √ nα (1 -α ) R ˜ X sc ij ,
Figure 4: FPR vs. FNR tradeoff. In both (a) and (b), we use π = 0 3 . , and thresholds ζ ∈ { 0 1 . , 0 2 . , . . . , 0 9 . } . We take p = 20000 for AMP, and p = 2000 for LP and CVX.
we have
$$Y _ { i ; } & = \sum _ { j = 1 } ^ { p } X _ { i j } ^ { s c } B _ { j, } + \Psi _ { i, ; } = \sum _ { j = 1 } ^ { p } \left ( \alpha W _ { r ( i ) \epsilon ( j ) } + \sqrt { \frac { n \alpha ( 1 - \alpha ) } { R } } \widetilde { X } _ { i j } ^ { s c } \right ) B _ { j, ; } + \Psi _ { i, ; } \\ & \implies Y _ { i, ; } - \sum _ { j = 1 } ^ { p } \alpha W _ { r ( i ) \epsilon ( j ) } B _ { j, ; } = \sqrt { \frac { n \alpha ( 1 - \alpha ) } { R } } \sum _ { j = 1 } ^ { p } \widetilde { X } _ { i j } ^ { s c } B _ { j, ; } + \Psi _ { i, ; } \,. \\ - \cdots & \cdots$$
Defining
$$\widetilde { Y } _ { i, \colon } \coloneqq \frac { 1 } { \sqrt { n \alpha ( 1 - \alpha ) / R } } \left ( Y _ { i, \colon } - \alpha \left ( W _ { r ( i ) 1 } \sum _ { j \in \mathcal { J } _ { 1 } } B _ { j, \colon } + \cdots + W _ { r ( i ) C } \sum _ { j \in \mathcal { J } _ { C } } B _ { j, \colon } \right ) \right ),$$
and ˜ Ψ i, : := 1 √ nα (1 -α / ) R Ψ i, : gives us the rescaled pooled data model:
$$\widetilde { Y } _ { i, ; } = B ^ { \top } \widetilde { X } _ { i, ; } ^ { s c } + \widetilde { \Psi } _ { i, ; } \in \mathbb { R } ^ { L }, \quad \text{for } i \in [ n ].$$
The sets ( J c ) c ∈ [ C ] are defined in (17). In the noiseless setting, the terms ∑ j ∈J c B j, : can be obtained with an extra C = O (1) tests, where the c th test only includes items from J c .
Model and noise scaling assumptions. The signal matrix B ∈ R p × L and the rescaled noise matrix ˜ Ψ ∈ R n × L Ψ are both independent of the design matrix. As p → ∞ , we assume that n/p → δ > 0 . As p, n → ∞ , the empirical distributions of the rows of B and ˜ Ψ each converge to well-defined limits. More precisely, B W → ¯ B and ˜ Ψ W → ¯ Ψ , for L -dimensional random vectors ¯ B ∼ Categorical ( π ) and ¯ Ψ ∼ P ¯ Ψ .
## 4.1 Matrix SC-AMP Algorithm
The goal is to recover B from ˜ Y generated according to the rescaled model (39). The matrix SCAMP algorithm is initialized with ̂ B 0 j, : = E [ ¯ ] B for j ∈ [ p ] , and ˜ Θ = 0 ˜ Y -˜ X B sc ̂ 0 . For iteration k ≥ 1 , we compute:
$$\widetilde { \Theta } ^ { k } & = \widetilde { Y } - \widetilde { X } ^ { s c } \widehat { B } ^ { k } + U ^ { k } \\ B ^ { k + 1 } & = V ^ { k } + \widehat { B } ^ { k } ; \quad \widehat { B } ^ { k + 1 } = f _ { k + 1 } ( B ^ { k + 1 }, \mathcal { C } ),$$
where U k ∈ R n × L , V k ∈ R p × L , and f k : R L +1 → R , which acts row-wise on its input, will be defined after the state evolution parameters are introduced. Note that the vector C ∈ R p is defined in (16).
State evolution. For k ≥ 0 , we iteratively compute the following L × L covariance matrices, starting from initialization ψ 0 c = Cov( ¯ ) B , c ∈ [ C ] :
$$\phi _ { r } ^ { k } = \text{Cov} [ \bar { \Psi } ] + \frac { 1 } { \delta _ { \text{in} } } \sum _ { r = 1 } ^ { C } \widetilde { W } _ { \text{rc} } \psi _ { \text{c} } ^ { k } \,, \quad r \in [ \text{R} ],$$
$$\phi _ { r } ^ { k } & = C o v [ \bar { \Psi } ] + \frac { 1 } { \delta _ { \text{in} } } \sum _ { r = 1 } ^ { \text{c} } \widetilde { W } _ { \text{rc} } \psi _ { \text{c} } ^ { k }, \quad \ \ r \in [ \text{R} ], \\ \text{T} _ { \text{c} } ^ { k } & = \left [ \sum _ { r = 1 } ^ { \text{R} } \widetilde { W } _ { \text{rc} } [ \phi _ { r } ^ { k } ] ^ { - 1 } \right ] ^ { - 1 }, \quad \ \psi _ { \text{c} } ^ { k + 1 } = \mathbb { E } \left [ \left ( f _ { k } ( \bar { B } + G _ { \text{c} } ^ { k }, \text{c} ) - \bar { B } \right ) \left ( f _ { k } ( \bar { B } + G _ { \text{c} } ^ { k }, \text{c} ) - \bar { B } \right ) ^ { \text{T} } \right ], \quad \text{c} \in [ \text{C} ]. \\ & \quad. \quad. \quad - \quad \ \sim$$
where G k c ∼ N (0 T ) , k c is independent of ¯ B , and we recall that ˜ W is defined in (18).
We define a matrix Q k ∈ R L R × L C in terms of sub-matrices Q k r c , ∈ R L × L , which for r ∈ [ R ] , c ∈ [ C ] , are given by
$$Q _ { r, \mathfrak { c } } ^ { k } = ( \phi _ { r } ^ { k } ) ^ { - 1 } \left ( \sum _ { r ^ { \prime } = 1 } ^ { \mathrm R } \widetilde { W } _ { r ^ { \prime } \mathfrak { c } } ( \phi _ { r ^ { \prime } } ^ { k } ) ^ { - 1 } \right ) ^ { - 1 }.$$
Then, the rows of the matrix U k ∈ R n × L in (40) are defined as
$$U _ { i, \colon } ^ { k } = \frac { 1 } { \delta _ { \text{in} } } \widetilde { \Theta } _ { i, \colon } ^ { k - 1 } \sum _ { c = 1 } ^ { c } \widetilde { W } _ { r ( i ), c } Q _ { r ( i ), c } ^ { k - 1 } \left \langle f _ { k } ^ { \prime } ( B _ { j, \colon } ^ { k }, c ) \right \rangle _ { c } ^ { \top }, \quad \text{for } i \in [ n ]$$
where ⟨·⟩ c is the empirical average over rows j ∈ J c , i.e., 〈 f ′ k ( B k j, : , c ) 〉 c = 1 p/ C ∑ j ∈J c f ′ k ( B k j, : , c ) . Next, the rows of the matrix V k ∈∈ R p × L are given by
$$V _ { j, \colon } ^ { k } = \sum _ { i = 1 } ^ { n } \widetilde { X } _ { i j } ^ { s c } \, \widetilde { \Theta } _ { i, \colon } ^ { k } Q _ { r ( i ), \mathfrak { c } ( j ) } ^ { k }, \quad \text{for } j \in [ p ].$$
Finally, the Bayes-optimal denoising function f k is given by the following, for j ∈ J c , c ∈ [ C ] :
$$f _ { k } ( B _ { j, \cdot } ^ { k }, \mathbf c ) = \mathbb { E } [ \bar { B } | \bar { B } + G _ { \mathbf c } ^ { k } = B _ { j, \cdot } ^ { k } ], \ \ G _ { \mathbf c } ^ { k } \sim \mathcal { N } \left ( 0, \mathbf T _ { \mathbf c } ^ { k } \right ) \text{ independent of } \bar { B }.$$
The matrix SC-AMP algorithm was proposed and analyzed in [58] for a model with a spatially coupled Gaussian design matrix. Similar to our analysis of SC-AMP for QGT, we could apply a
reduction technique along with the universality result of [36] to establish a state evolution characterization for the matrix SC-AMP applied to pooled data. Such a result would be analogous to Theorem 3.2, and show that for each iteration k ≥ 1 , the joint empirical distribution of the rows of ( B,B k ) converges as:
$$( B, B ^ { k } ) _ { \mathcal { J } _ { c } ; } \stackrel { W _ { 2 } } { \to } ( \bar { B }, \bar { B } + G _ { c } ^ { k } ), \ \ c \in [ C ].$$
To analyze the limiting MSE and error rate of the matrix SC-AMP algorithm, we need to characterize the fixed point of the state evolution recursion in (41) (as k →∞ ). In QGT, the state evolution fixed point was characterized via the minimizer of a scalar potential function (Theorem 3.5). Extending this approach to the pooled data setting is challenging as the state evolution parameters are now L × L matrices rather than scalars. In the following section, we circumvent this issue by showing that a suboptimal AMP algorithm still achieves almost-exact recovery for any δ > 0 . The suboptimal algorithm applies the SC-AMP algorithm column-wise to ˜ Y ∈ R n × L , ignoring the correlation between the columns of the signal matrix B .
## 4.2 Almost-Exact Recovery via Column-wise SC-AMP
Given ˜ X ,Y sc ˜ from the rescaled model (39), we run the SC-AMP algorithm column-wise on ˜ Y . Specifically, for l ∈ [ L ] , run the SC-AMP algorithm (15) with inputs Y : ,l and ˜ X sc to produce the estimate ̂ B k : ,l after k ≥ 1 iterations. For the SC-AMP algorithm applied to column l ∈ [ L ] , the denoiser f k in (19) is computed with ¯ β ∼ Bernoulli ( π l ) . (We recall that the rows of the signal follow the prior ¯ B ∼ Categorical ( π ) where π = ( π , . . . , π 1 L ) .)
The column-wise SC-AMP algorithm can be viewed as an instance of the matrix SC-AMP algorithm with a suboptimal denoiser, obtained by replacing the conditional expectation E [ ¯ B | ¯ + B G k c ] in (43) with the marginal conditional expectations E [ ¯ B l | ¯ B l +( G k c ) ] l , for l ∈ [ L ] . Let us define the quantized estimate after k iterations of the column-wise SC-AMP algorithm to be
$$\widetilde { B } _ { j l } ^ { k } = \begin{cases} 1 & \text{if } \widehat { B } _ { j l } ^ { k } > 0. 5 \\ 0 & \text{otherwise}, \end{cases}$$
where ̂ B k is the estimate obtained from the SC-AMP algorithm. In practice, we can quantize the estimate in a better manner, by setting the largest entry in the row of ̂ B to one and the remaining entries in the row to zero. We do not use this form of quantization for our almost-exact recovery result since we want to directly apply the SC-AMP results for QGT to the pooled data setting.
Theorem 4.1. Consider the noiseless pooled data problem with the assumptions stated on p.17, for any δ > 0 . There exist finite ω 0 and k 0 such that for all ω > ω 0 , k > k 0 , and sufficiently large Λ , the quantized estimate ˜ B k ∈ { 0 1 , } p × L produced by the column-wise SC-AMP algorithm almost surely satisfies:
̸
$$\lim _ { k \to \infty } \lim _ { p \to \infty } \frac { 1 } { p } \sum _ { j = 1 } ^ { p } 1 \left \{ \widetilde { B } _ { j, \colon } ^ { k } \neq B _ { j, \colon } \right \} = 0.$$
̸
̸
Proof. The model assumptions imply that for each l ∈ [ L ] , the empirical distribution of column B : ,l converges in Wasserstein distance to Bernoulli ( π l ) . By Corollary 3.8, we have that the SC-AMP algorithm applied to ˜ Y : ,l satisfies lim k →∞ lim p →∞ 1 p ∑ p j =1 1 { ˜ B k j,l = B j,l } = 0 almost surely, for each l ∈ [ L ] . The result follows by noting that 1 p ∑ p j =1 1 { ˜ B k j, : = B j, : } ≤ 1 p ∑ L l =1 ∑ p j =1 1 { ˜ B k jl = B jl } .
̸
(a) SC-AMP vs. iid-AMP
Figure 5: Noiseless Pooled Data with π = [1 3 1 / , / , 3 1 / 3] with p = 20000 and spatial coupling parameters ω = 6 Λ = 40 , . In (b), we set p = 1000 for LP due to the high computational cost. Error bars indicate one standard deviation.
We can also obtain error guarantees in the low-noise regime for the column-wise SC-AMP algorithm, similar to Theorem 3.7 and Corollary 3.8. We remark that although column-wise SC-AMP is convenient for theoretical analysis, at finite dimensions it is inferior to the matrix SC-AMP algorithm that takes advantage of the correlation in the columns of B via the denoiser in (43). This is illustrated in the numerical experiments below.
## 4.3 Numerical Simulations
We present simulation results for noiseless pooled data using matrix SC-AMP, measuring the performance via the normalized correlation 1 p ∑ p j =1 ⟨ ̂ B k j, : , B j, : ⟩ after k iterations of the algorithm. By (44), we this normalized correlation converges as
$$\frac { 1 } { p } \sum _ { j = 1 } ^ { p } \langle \widehat { B } _ { j, \cdot } ^ { k }, B _ { j, \cdot } \rangle \stackrel { a. s. } { \to } \frac { 1 } { \text{C} } \sum _ { \text{c=1} } ^ { \text{c} } \mathbb { E } \left [ \langle f _ { k } ( \bar { B } + G _ { \text{c} } ^ { k }, \text{c} ), \bar { B } \rangle \right ].$$
Each point on the AMP performance curves is obtained from 10 independent runs, where in each run, the algorithm is executed for 300 iterations. Our benchmark will be the linear programming (LP) estimator tailored to the pooled data problem [14, Section 4.1]. Recall from (4) that the information theoretic lower bound on the number of tests for the noiseless pooled data problem is
$$n ^ { * } = \frac { p } { \log p } \left ( \max _ { r \in \{ 1, \dots, L - 1 \} } \frac { 2 [ H ( \pi ) - H ( \pi ^ { ( r ) } ) ] } { L - r } \right ).$$
Figure 5 shows how the normalized correlation varies with the sampling ratio δ for pooled data with L = 3 equally likely categories. Figure 5a shows that the state evolution prediction for performance of the matrix SC-AMP algorithm (SC-SE curves) improves as the spatial coupling
Figure 6: Matrix-SC AMP vs column-wise SC-AMP for noiseless pooled data with π = [1 / , 3 1 / , 3 1 / 3] , p = 20000 and spatial coupling parameters ω = 6 Λ = 40 , .
parameters ( ω, Λ) increase from (6 40) , to (20 200) , . As in QGT, the gap between the empirical performance of matrix SC-AMP and the state evolution prediction for (6 , 40) is due to finite length effects. We did not implement the matrix SC-AMP algorithm for (20 200) , as it requires a large amount of computational memory. In Figure 5b, we observe that iid-AMP, iid-LP, and SC-LP have similar performance, and SC-AMP outperforms all of these, justifying the use of a SC design matrix with matrix SC-AMP for recovery. We also implemented the iterative hard thresholding algorithm [14, Section 4.1] but found that it performed significantly worse AMP and LP, and so omitted it from our comparisons.
Figure 6 compares the performance of matrix SC-AMP with the column-wise SC-AMP algorithm. To make the algorithms comparable, the estimates from each algorithm (and the corresponding SE) were quantized in the same way after their final iteration, using the rule in (45). This leads to a slight difference in the AMP performance curves and the theoretical SE estimates compared to Figure 5, where no quantization was used. As expected, the matrix SC-AMP algorithm outperforms the column-wise SC-AMP algorithm since the former takes advantage of the correlation within each row of the matrix signal. Nevertheless, the column-wise SC-AMP algorithm performs slightly better than the matrix AMP algorithm with an i.i.d. matrix.
## 5 Proof of Theorem 3.2
We start by defining generalized white noise matrices , which will be used in the proof of the theorem.
Definition 5.1. [36, Definition 2.15] A generalized white noise matrix ˜ X ∈ R n × p with a (deterministic) variance profile S ∈ R n × p is one satisfying the following conditions, for i ∈ [ n , j ] ∈ [ p ] :
- 1. All entries ˜ X ij are independent.
- 2. Each entry ˜ X ij has mean 0, variance n -1 S ij , and higher moments satisfying, for each integer
m ≥ 3 ,
$$\lim _ { n, p \to \infty } p \cdot \max _ { i \in [ n ] } \max _ { j \in [ p ] } \mathbb { E } \left [ | \widetilde { X } _ { i j } | ^ { m } \right ] = 0.$$
## 3. For a constant C > 0 ,
$$\max _ { i \in [ n ] } \max _ { j \in [ p ] } S _ { i j } \leq C, \quad \lim _ { n, p \to \infty } \max _ { i \in [ n ] } \left | \frac { 1 } { p } \sum _ { j = 1 } ^ { p } S _ { i j } - 1 \right | = 0, \quad \lim _ { n, p \to \infty } \max _ { j \in [ p ] } \left | \frac { 1 } { n } \sum _ { i = 1 } ^ { n } S _ { i j } - 1 \right | = 0. \quad ( 4 7 )$$
Definition 5.1 simplifies for the case of S ij = 1 for all ( i, j ) ∈ [ n ] × [ p ] . In this case, the entries are all i.i.d. with variance 1 /n , the third condition in the definition is trivially satisfied, and the second condition requires moments of order 3 and higher to decay faster than 1 /p . The rescaled i.i.d. design matrix ˜ X iid in (9) is a generalized white noise matrix, but the rescaled spatially coupled matrix ˜ X sc in (10) is not. Indeed, ˜ X sc satisfies the first two requirements in Definition 5.1 and from (11), its variance profile of satisfies the first and last conditions in (47), but not the second: for i ∈ [ n ] , we have 1 p ∑ p j =1 S ij = R C , which is close to, but not equal to 1 for large Λ /ω .
We prove Theorem 3.2 via a more general result, for a generalized linear model with a spatial coupled design, where the observations y i ∈ R are generated as:
$$\tilde { y } _ { i } = q \left ( \left ( \widetilde { X } _ { i, \cdot } ^ { \text{sc} } \right ) ^ { \top } \beta, \, \widetilde { \Psi } _ { i } \right ) = q \left ( \Theta _ { i }, \widetilde { \Psi } _ { i } \right ), \quad \text{for } i \in [ n ],$$
Here β ∈ R p is the signal to be estimated, ˜ Ψ ∈ R n is a noise vector, and q : R 2 → R is a known output function. We also allow ˜ X sc to be more general than the one in Definition 3.1. The generalized spatially coupled matrix ˜ X sc ∈ R n × p consists of independent zero-mean entries whose variances are specified by a generic base matrix ˜ W ∈ R R × C , which satisfies the following conditions:
$$\sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } = 1 \quad \text{for $c\in [C]$,} \quad \text{and} \quad \kappa _ { 1 } \leq \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \leq \kappa _ { 2 }, \quad \text{for $r\in [R]$,}$$
for some κ , κ 1 2 > 0 . Given a base matrix ˜ W satisfying (49), we construct the spatially coupled matrix ˜ X sc with independent entries drawn from a distribution with zero-mean and variance E [ | ˜ X sc ij | 2 ] = ˜ W r ( ) i c ( j ) n/R , for i ∈ [ n , j ] ∈ [ p ] . We also assume that the higher moments E [ | ˜ X sc ij | m ] for m ≥ 3 , satisfy (46).
The first condition in (49) ensures that the expected squared norm of each column of ˜ X sc is 1 , and the second condition in (49) bounds the variance of each entry of ˜ X β sc from above and below.
## High-level sketch of proof of Theorem 3.2. The proof consists of three reductions.
- 1. In Section 5.1, we introduce the spatially coupled generalized approximate message passing algorithm (SC-GAMP) for the generalized linear model in (48), and characterize its performance via state evolution (Theorem 5.2). We then reduce the SC-AMP algorithm in (15) to SC-GAMP, and use the state evolution result of the latter to prove Theorem 3.2.
- 2. To prove the state evolution result for SC-GAMP (Theorem 5.2), we show that the algorithm can be written as an instance of an abstract matrix-AMP iteration defined via a generalized white noise matrix. This reduction, shown in Appendix A.1, is similar to the one used in [35, Appendix A] for reducing the SC-GAMP algorithm for a Gaussian design to an abstract matrix-AMP iteration.
- 3. To prove the state evolution result for the abstract matrix-AMP (Theorem A.1), we show that it is a special case of an AMP iteration for generalized white noise matrices, for which a rigorous state evolution result was established in [36]. We refer to the latter iteration as UAMP, where the 'U' stands for universal. The technique for reducing the abstract matrix-AMP to U-AMP is similar to the one presented in [14]. This is shown in Appendix A.2.
As before, to simplify notation, for vectors a ∈ R p and b ∈ R n , we will use a c := a J c and b r := b I r , where J c and I r are defined in (17). There will be no notation simplification for matrices.
## 5.1 The SC-GAMP Algorithm and its State Evolution
The SC-GAMP algorithm aims to estimate the signal β ∈ R p from observations ˜ y ∈ R n generated according to the generalized linear model (48). For iteration k ≥ 0 , the algorithm computes:
$$& \Theta ^ { k } = \widetilde { X } ^ { \text{sc} } \hat { \beta } ^ { k } - b ^ { k } \odot \widehat { R } ^ { k - 1 } ; \quad \widehat { R } ^ { k } = g _ { k } ( \Theta ^ { k }, \tilde { y }, \mathcal { R } ) \\ & \beta ^ { k + 1 } = ( \widetilde { X } ^ { \text{sc} } ) ^ { \top } \widehat { R } ^ { k } - \epsilon ^ { k } \odot \hat { \beta } ^ { k } ; \quad \hat { \beta } ^ { k + 1 } = f _ { k + 1 } ( \beta ^ { k + 1 }, \mathcal { C } ),$$
where ⊙ denotes element-wise product. The algorithm is initialized with some ˆ β 0 ∈ R p and Θ = 0 ˜ X β sc ˆ 0 . The functions g k and f k +1 act row-wise on their input, and
$$\mathcal { C } & = ( \underbrace { 1, \dots, 1 } _ { p / C \text{ entries } } _ { p / C \text{ entries } }, \underbrace { 2, \dots, 2 } _ { p / C \text{ entries } }, \dots, \underbrace { C, \dots, C } _ { p / C \text{ entries } } ) ^ { \top } \in \mathbb { R } ^ { p }, \\ \mathcal { R } & = ( \underbrace { 1, \dots, 1 } _ { n / R \text{ entries } }, \underbrace { 2, \dots, 2 } _ { n / R \text{ entries } }, \dots, \underbrace { R, \dots, R } _ { n / R \text{ entries } } ) ^ { \top } \in \mathbb { R } ^ { n }.$$
The entries of c k ∈ R p and b k ∈ R n are defined as follows, for j ∈ [ p ] , i ∈ [ n ] :
$$c _ { j } ^ { k } = \sum _ { r = 1 } ^ { \mathrm R } \frac { \widetilde { W } _ { \mathrm c } } { n / \mathrm R } \sum _ { i \in \mathcal { I } _ { r } } \partial _ { 1 } g _ { k } ( \Theta _ { i } ^ { k }, \tilde { y } _ { i }, r ), \ \ b _ { i } ^ { k } = \sum _ { c = 1 } ^ { \mathrm c } \frac { \widetilde { W } _ { \mathrm c } } { n / \mathrm R } \sum _ { j \in \mathcal { J } _ { c } } \partial _ { 1 } f _ { k } ( \beta _ { j } ^ { k }, \mathrm c ).$$
State evolution. The 'memory' terms -b k ⊙ ̂ R k -1 and -c k ⊙ ˆ β k in (50) debias the iterates Θ k and β k +1 , ensuring that their empirical distributions are accurately captured by state evolution in the high-dimensional limit. Theorem 5.2 below shows that for each k ≥ 1 and c ∈ [ C ] , the empirical distribution of β k c converges to the distribution of µ k β, c ¯ + β G k β, c where ¯ β is the random variable representing the limiting distribution of the entries of the signal β c , and G k β, c ∼ N ( 0 ( , σ β, c ) 2 ) is independent of ¯ β . The deterministic parameters µ k β, c ∈ R and σ k β, c ∈ R are defined below. The result implies that the empirical distribution of the estimate ˆ β k c converges to the distribution of f k ( µ k β. c ¯ + β G k β, c ) . Thus, f k can be viewed as a denoising function that can be tailored to take
advantage of the prior in ¯ β . Theorem 5.2 also shows that the joint empirical distribution of the rows of (Θ r , Θ ) k r converges to N (0 Σ , k, r ) , where Σ k, r ∈ R 2 × 2 is defined below.
We now describe the state evolution recursion defining µ k β, c , σ k β, c ∈ R and Σ k, r ∈ R 2 × 2 . Define ¯ g k : R 3 × [ R ] → R such that
$$g _ { k } ( \Theta _ { i } ^ { k }, \tilde { y } _ { i }, r ) = \bar { g } _ { k } ( \Theta _ { i }, \Theta _ { i } ^ { k }, \widetilde { \Psi } _ { i }, r ) \text{ for } i \in \mathcal { I } _ { r },$$
since ˜ y i = q (Θ i , ˜ Ψ ) i . Starting with an initializer Σ 0 , r ∈ R 2 × 2 for r ∈ [ R ] (defined later in (56)), the state evolution parameters are iteratively computed as follows for k ≥ 0 , and r ∈ [ R ] , c ∈ [ C ] :
$$\mu _ { \beta, c } ^ { k + 1 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r e } \mathbb { E } [ \partial _ { 1 } \bar { g } _ { k } ( Z _ { r }, Z _ { r } ^ { k }, \bar { \Psi }, r ) ], \quad ( \sigma _ { \beta, c } ^ { k + 1 } ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r e } \mathbb { E } [ \bar { g } _ { k } ( Z _ { r }, Z _ { r } ^ { k }, \bar { \Psi }, r ) ^ { 2 } ], \\ \Sigma ^ { k + 1, r } = \begin{bmatrix} \Sigma _ { 1 1 } ^ { k + 1, r } & \Sigma _ { 1 2 } ^ { k + 1, r } \\ \Sigma _ { 2 1 } ^ { k + 1, r } & \Sigma _ { 2 2 } ^ { k + 1, r } \end{bmatrix}, \\. \quad \, \quad \, \, \,$$
where ( Z , Z r k r ) ∼ N (0 Σ , k, r ) are independent of ¯ Ψ , and
$$Z _ { r }, Z _ { r } ^ { \kappa } ) & \sim \mathcal { N } ( 0, \Sigma ^ { \kappa, r } ) \text{ are independent of } \Psi, \text{ and} \\ \Sigma _ { 1 1 } ^ { k + 1, r } & = \mathbb { E } [ ( Z _ { r } ) ^ { 2 } ] = \frac { \mathbb { E } [ \bar { \beta } ^ { 2 } ] } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \\ \Sigma _ { 1 2 } ^ { k + 1, r } & = \Sigma _ { 2 1 } ^ { k + 1, r } = \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \mathbb { E } [ \bar { \beta } f _ { k + 1 } ( \mu _ { \beta, c } ^ { k + 1 } \bar { \beta } + G _ { \beta, c } ^ { k + 1 }, c ) ] \\ \Sigma _ { 2 2 } ^ { k + 1, r } & = \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \mathbb { E } [ f _ { k + 1 } ( \mu _ { \beta, c } ^ { k + 1 } \bar { \beta } + G _ { \beta, c } ^ { k + 1 }, c ) ^ { 2 } ]. \\ ^ { k + 1 } _ { a _ { \kappa } ^ { \kappa } } & \sim \mathcal { N } ( 0, ( \sigma _ { a _ { \kappa } ^ { k + 1 } } ) ^ { 2 } ) \text{ is independent of } \bar { \beta }.$$
Here G k +1 β, c ∼ N (0 ( , σ k +1 2 β, c ) ) is independent of ¯ β .
The SC-GAMP algorithm and its state evolution equations are similar to those introduced in [35], the only difference being that the SC design matrix ˜ X sc ∈ R n × p is now a generalized spatially coupled matrix instead of the spatially coupled Gaussian one used in [35]. We note that ˜ X sc is not a generalized white noise matrix since it has variance profile S sc ij := R ˜ W r ( ) i , c ( j ) , which is not guaranteed to satisfy the condition max i ∈ [ n ] ∣ ∣ 1 p ∑ p j =1 S sc ij -1 ∣ ∣ → 0 in Definition 5.1. Nevertheless, ˜ X sc can be related to a generalized white noise matrix ˜ X ∈ R n × p defined as follows. For i ∈ [ n , j ] ∈ [ p ] , let:
̸
$$\widetilde { X } _ { i j } \coloneqq \begin{cases} \frac { \widetilde { X } _ { i j } ^ { s c } } { \sqrt { R \cdot \widetilde { W } _ { r ( i ), c ( j ) } } }, & \text{ if } \widetilde { W } _ { r ( i ), c ( j ) } \neq 0, \\ \underset { \text{indep.} } { \sim } \mathcal { N } ( 0, \frac { 1 } { n } ), & \text{ otherwise.} \end{cases}$$
(In the second line of the definition, we could use any sub-Gaussian distribution with mean zero and variance 1 /n instead of N (0 1 , /n ) .) From the construction of ˜ X sc (see below (49)), it follows that ˜ X is a generalized white noise matrix with variance profile S ij = 1 for all ( i, j ) ∈ [ n ] × [ p ] .
The state evolution result for SC-GAMP requires the following assumptions on the model and the algorithm:
(A1) As n, p → ∞ , we have n p → δ . The signal β , initializer ˆ β 0 , and the noise vector ˜ Ψ are independent of X sc , and their empirical distributions have well-defined limits. There exist random variables ¯ β ∼ P ¯ β and ¯ Ψ ∼ P ¯ Ψ with β W → ¯ β and ˜ Ψ W → ¯ Ψ , respectively.
- (A2) As p → ∞ , ( β , β c ˆ ) 0 c W → ( ¯ β, β ¯ ) 0 c almost surely, for joint laws ( ¯ β, β ¯ ) 0 c having finite moments of all orders, for c ∈ [ C ] . Furthermore, multivariate polynomials are dense in the real L 2 -spaces of functions f : R → R and g : R 2 → R with the inner-products
$$\langle f, \tilde { f } \rangle \coloneqq \mathbb { E } [ f ( \bar { \Psi } ) \tilde { f } ( \bar { \Psi } ) ] \ \text{ and } \ \langle g, \tilde { g } \rangle \coloneqq \mathbb { E } [ g ( \bar { \beta }, \bar { \beta } _ { \mathfrak { c } } ^ { 0 } ) \tilde { g } ( \bar { \beta }, \bar { \beta } _ { \mathfrak { c } } ^ { 0 } ) ].$$
- (A3) For k ≥ 0 and r ∈ [ R ] , c ∈ [ C ] , the functions f k ( · , c ) and ¯ ( g k · , · , · , r ) are each continuous, Lipschitz w.r.t. the first argument, and satisfy the polynomial growth condition in (5) for some order r ≥ 1 .
- (A4) The matrix ˜ X defined as in (55) satisfies ∥ ˜ X ∥ op < C for some constant C , and for any fixed polynomial function f † : R 2 R +2 → R , as n, p →∞ ,
$$\max _ { i \in [ n ] } \Big | & \Big < f ^ { \dagger } \Big ( \beta _ { \epsilon }, \beta _ { \epsilon } \sqrt { \widetilde { R W } _ { 1 } }, \dots, \beta _ { \epsilon } \sqrt { \widetilde { R W } _ { \epsilon } }, \hat { \beta } _ { \epsilon } ^ { 0 } \sqrt { \widetilde { R W } _ { 1 } }, \dots, \hat { \beta } _ { \epsilon } ^ { 0 } \sqrt { \widetilde { R W } _ { \epsilon } }, c \Big ) \odot S _ { i, \mathcal { J } } \Big > \\ & - \Big < f ^ { \dagger } \Big ( \beta _ { \epsilon }, \beta _ { \epsilon } \sqrt { \widetilde { R W } _ { 1 } }, \dots, \beta _ { \epsilon } \sqrt { \widetilde { R W } _ { \epsilon } }, \hat { \beta } _ { \epsilon } ^ { 0 } \sqrt { \widetilde { R W } _ { 1 } }, \dots, \hat { \beta } _ { \epsilon } ^ { 0 } \sqrt { \widetilde { R W } _ { \epsilon } }, c \Big ) \Big > \cdot \langle S _ { i, \mathcal { J } } \rangle \Big | \stackrel { a. s } { \to } 0,$$
for all c ∈ [ C ] , where S is the variance profile of ˜ X (see Definition 5.1) and f † acts element-wise on β c .
- (A5) For any fixed polynomial function f ‡ : R 2 → R , as n, p →∞ ,
$$\max _ { j \in [ p ] } \left | \left \langle f ^ { \dagger } \left ( \widetilde { \Psi } _ { r }, r \right ) \odot S _ { \mathcal { I } _ { r }, j } \right \rangle - \left \langle f ^ { \dagger } \left ( \widetilde { \Psi } _ { r }, r \right ) \right \rangle \cdot \left \langle S _ { \mathcal { I } _ { r }, j } \right \rangle \right | \stackrel { a. s. } { \to } 0,$$
for all r ∈ [ R ] , where f ‡ acts element-wise on ˜ Ψ r .
The state evolution recursion in (53)-(54) is initialized with
$$\Sigma ^ { 0, r } = \left [ \begin{smallmatrix} \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \mathbb { E } [ \bar { \beta } ^ { 2 } ] \\ \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \mathbb { E } [ \bar { \beta } ^ { 0 } \bar { \beta } ] \\ \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \mathbb { E } [ ( \bar { \beta } ^ { 0 } _ { c } ) ^ { 2 } ] \end{smallmatrix} \right ], \quad r \in [ R ].$$
Theorem 5.2 (State evolution for SC-GAMP) . Consider the GLM in (48) with spatially coupled design ˜ X sc defined via a base matrix satisfying (49) , and signal estimation using the SC-GAMP recursion in (50) . Let Assumptions (A1)-(A5) be satisfied, and assume σ 1 β, c > 0 for c ∈ [ C ] . Then for each k ≥ 0 , we have
$$\left ( \beta _ { \epsilon }, \beta _ { \epsilon } ^ { k + 1 } \right ) \stackrel { W _ { 2 } } { \to } \left ( \bar { \beta }, \mu _ { \beta, \epsilon } ^ { k + 1 } \bar { \beta } + G _ { \beta, \epsilon } ^ { k + 1 } \right ), \ \left ( \Theta _ { r }, \Theta _ { r } ^ { k } \right ) \stackrel { W _ { 2 } } { \to } \left ( Z _ { r }, Z _ { r } ^ { k } \right ),$$
almost surely as n, p →∞ with n/p → δ .
The proof is given in Appendix A.1. We now use Theorem 5.2 to prove Theorem 3.2.
## 5.2 Proof of Theorem 3.2 using Theorem 5.2
We first verify that rescaled QGT model
$$\tilde { y } _ { i } = q \left ( \left ( \widetilde { X } _ { i, \colon } ^ { s c } \right ) ^ { \top } \beta, \, \widetilde { \Psi } _ { i } \right ) = \left ( \widetilde { X } _ { i, \colon } ^ { s c } \right ) ^ { \top } \beta + \widetilde { \Psi } _ { i }$$
is a special case of the GLM (48) with a generalized spatially coupled design ˜ X sc constructed as described below (49). The rescaled QGT design ˜ X sc in (10) has independent zero mean entries with variances ˜ W r ( ) i c ( j ) n/R for i ∈ [ n , j ] ∈ [ p ] , where from (18) we have ˜ W rc = 1 /ω for c ≤ r ≤ c + ω -1 , and 0 otherwise. Moreover, this ˜ W satisfies the conditions in (49).
Next, we show that the SC-AMP in (15) is a special case of the SC-GAMP algorithm by choosing
$$f _ { k } ( \beta _ { j } ^ { k } \,, c ) = \mathbb { E } [ \bar { \beta } \, | \, \mu _ { \beta, c } ^ { k } \bar { \beta } + G _ { \beta, c } ^ { k } = \beta _ { j } ^ { k } ], \ \ j \in \mathcal { J } _ { c }.$$
$$g _ { k } ( \Theta _ { i } ^ { k }, \tilde { y } _ { i }, r ) = \frac { \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \cdot \colon } { \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } }, \quad \text{for $i\in \mathcal{I}_{r}$.}$$
The choices in (58) and (59) are based on the Bayes-optimal denoisers for an i.i.d. design and Gaussian noise ¯ Ψ ∼ N (0 , σ 2 ) (see [31, Section 4.2]), and take into account the block-wise dependence of the state evolution parameters. With this choice of f k , in (54) we have E [ ¯ βf k ] = E [ f 2 k ] which implies that Σ k, r 12 = Σ k, r 21 = Σ k, r 22 , for k ≥ 1 .
With our choice of denoisers, the iterate ̂ R k ∈ R n in (50) can be written as
$$\widehat { R } ^ { k } = Q ^ { k } \odot ( \tilde { y } - \Theta ^ { k } ),$$
where the entries of Q k ∈ R n are Q k i = Σ ( k, r ( ) i 11 -Σ k, r ( ) i 12 + σ 2 ) -1 , for i ∈ [ n ] . For j ∈ [ p ] , we have:
$$c _ { j } ^ { k } & = \sum _ { r = 1 } ^ { R } \frac { \widetilde { W } _ { \text{rc} } } { n / R } \sum _ { i \in \mathcal { I } _ { r } } \partial _ { 1 } g _ { k } ( \Theta _ { i } ^ { k }, \tilde { \jmath } _ { y }, r ) = \frac { R } { n } \sum _ { r = 1 } ^ { R } \widetilde { W } _ { \text{rc} } \sum _ { i \in \mathcal { I } _ { r } } \frac { - 1 } { \Sigma _ { 1 1 } ^ { k, r ( i ) } - \Sigma _ { 1 2 } ^ { k, r ( i ) } + \sigma ^ { 2 } } \\ & = \frac { R } { n } \sum _ { r = 1 } ^ { R } \widetilde { W } _ { \text{rc} } \cdot \frac { n } { R } \cdot \frac { - 1 } { \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } } = - \sum _ { r = 1 } ^ { R } \widetilde { W } _ { \text{rc} } \left ( \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } \right ) ^ { - 1 }.$$
Next, we have
$$( \sigma _ { \beta, c } ^ { k + 1 } ) ^ { 2 } & = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \mathbb { E } [ \bar { g } _ { k } ( Z _ { r }, Z _ { r } ^ { k }, \bar { \Psi }, r ) ^ { 2 } ] \stackrel { ( a ) } { = } \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \mathbb { E } \left [ \frac { ( Z _ { r } + \bar { \Psi } - Z _ { r } ^ { k } ) ^ { 2 } } { ( \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } ) ^ { 2 } } \right ] \\ & \stackrel { ( b ) } { = } \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \frac { \sigma ^ { 2 } + \mathbb { E } [ ( Z _ { r } - Z _ { r } ^ { k } ) ^ { 2 } ] } { ( \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } ) ^ { 2 } } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } ( \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } ) ^ { - 1 } \stackrel { ( c ) } { = } - c _ { j } ^ { k }, \\. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad. \quad.$$
where (a) applies (59), (b) uses the independence between ( Z , Z r k r ) and ¯ Ψ , and (c) uses (61). Substituting the definitions of Σ k, r 11 and Σ k, r 12 in (54) into (62), we get
$$( \sigma _ { \beta, \epsilon } ^ { k + 1 } ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r \epsilon } \left ( \sigma ^ { 2 } + \frac { 1 } { \delta _ { \ln } } \sum _ { \epsilon ^ { \prime } = 1 } ^ { C } \widetilde { W } _ { r \epsilon ^ { \prime } } \mathbb { E } \left [ ( \bar { \beta } - f _ { k } ( \mu _ { \beta, \epsilon } ^ { k } \bar { \beta } + G _ { \beta, \epsilon } ^ { k }, \epsilon ) ) ^ { 2 } \right ] \right ) ^ { - 1 }.$$
We also have the identity
$$\mu _ { \beta, \epsilon } ^ { k + 1 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { \text{rc} } \mathbb { E } [ \partial _ { 1 } \bar { g } _ { k } ( Z, Z _ { r } ^ { k }, \bar { \Psi }, r ) ] \stackrel { ( a ) } { = } \sum _ { r = 1 } ^ { R } \widetilde { W } _ { \text{rc} } ( \Sigma _ { 1 1 } ^ { k, r } - \Sigma _ { 1 2 } ^ { k, r } + \sigma ^ { 2 } ) ^ { - 1 } \stackrel { ( b ) } { = } ( \sigma _ { \beta, \epsilon } ^ { k + 1 } ) ^ { 2 },$$
where (a) uses (59) and ¯ y r = Z r + ¯ Ψ , and (b) uses the last equality in (62).
Letting χ k c = σ k β, c , we observe that the update equations in (63)-(64) match the state evolution recursion of the SC-AMP algorithm in (21). Then, substituting (58)-(63) into SC-GAMP in (50), followed by a change of variables from ˜ Θ k to ˜ Θ := ˜ k y -Θ k , gives us the SC-AMP algorithm in (15). Finally, we check that the assumptions of Theorem 5.2 are satisfied:
- · Assumptions (A1) and (A2) hold due to the model assumptions in Section 2, the noise scaling assumption (p. 9), and the SC-AMP initialization ˆ β 0 = E [ ¯ ]1 β p . Recalling that ¯ β ∼ Bernoulli ( π ) for the QGT model, the state evolution initialization in (56) becomes
$$\Sigma ^ { 0, r } = \frac { 1 } { \delta _ { \text{in} } } \left [ \begin{matrix} \sum _ { c = 1 } ^ { \text{C} } \widetilde { W } _ { r c } \pi & \sum _ { c = 1 } ^ { \text{C} } \widetilde { W } _ { r c } \pi ^ { 2 } \\ \sum _ { c = 1 } ^ { \text{C} } \widetilde { W } _ { r c } \pi ^ { 2 } & \sum _ { c = 1 } ^ { \text{C} } \widetilde { W } _ { r c } \pi ^ { 2 } \end{matrix} \right ], \quad r \in [ R ].$$
Using this in (63), we obtain that σ 1 β, c = χ 1 c , where the latter is defined in (22).
- · (A3) . With ¯ β ∼ Bernoulli ( π ) , the denoiser f k ( · , c ) in (58) can be explicitly computed (see (116)), and the choice for ¯ ( g k · , · , · , r ) is given by (59) and (52). From these expressions, it can be verified that both functions are continuous, Lipschitz w.r.t. the first argument, and satisfy the polynomial growth condition with r = 2 .
- · (A4) and (A5) . Recalling the definition of ˜ X in (55) and of ˜ X sc in (10), we note the matrix √ nX ˜ has independent sub-Gaussian entries of variance 1. Using a concentration inequality for the operator norm of sub-Gaussian matrices [59, Theorem 4.4.5] together with the BorelCantelli lemma, we obtain that ∥ ˜ X ∥ op < C almost surely for sufficiently large p . Since the variance profile S ij = 1 for all ( i, j ) , the second condition in (A3) is trivially satisfied. Assumption (A5) is similarly satisfied.
This completes the proof.
## 6 Proof of Theorem 3.5
## 6.1 Proof of (32)
The idea is to rewrite the SC-AMP state evolution in (21) in terms of a general coupled recursion analyzed by Yedla et al. in [37]. We then apply the fixed point characterization of [37, Theorem 1] to the SC-AMP state evolution to obtain (32).
General coupled recursion [37]. Let X = [0 , x max ] , Y = [0 , y max ] with x max , y max ∈ (0 , ∞ ) . Let f : Y → X be a non-decreasing C 1 function, and let g : X → Y be a strictly increasing C 2 function with y max = g x ( max ) . (We say a function f : Z → R is C d if its d th derivative exists and
is continuous on Z .) Consider a matrix A ∈ R C × R with R = C + ω -1 , whose entries are as defined as follows, for r ∈ [ R ] , c ∈ [ C ] :
$$A _ { \text{cr} } = \begin{cases} \frac { 1 } { \omega } & \text{if $c\leq r\leq c+\omega-1$} \\ 0 & \text{otherwise.} \end{cases}$$
Using A , we define the following coupled recursion. For r ∈ [ R ] :
$$y _ { r } ^ { k + 1 } = g ( x _ { r } ^ { k } ), \ \ x _ { r } ^ { k + 1 } = \sum _ { c = 1 } ^ { C } A _ { c r } f \left ( \sum _ { r ^ { \prime } = 1 } ^ { R } A _ { c r ^ { \prime } } y _ { r ^ { \prime } } ^ { k + 1 } \right ),$$
The recursion is initialized with x 0 r = x max for r ∈ [ R ] . This initialization, along with the monotonicity of f and g , ensures that the coupled recursion converges to a fixed point [37]. The fixed point { lim k →∞ x k r } r ∈ [ R ] is characterized by the lemma below in terms of the following potential function:
$$V ( x ) \coloneqq x g ( x ) - \int _ { 0 } ^ { x } g ( z ) d z - \int _ { 0 } ^ { g ( x ) } f ( z ) d z.$$
Lemma 6.1. [37, Theorem 1] For any γ > 0 , there exists ω 0 < ∞ such that for all ω > ω 0 and C ∈ [1 , ∞ ] , the fixed point x ∞ r := lim k →∞ x k r , for r ∈ [ R ] , of the coupled recursion in (65) satisfies the upper bound
$$\max _ { r \in [ R ] } x _ { r } ^ { \infty } \leq \max \left \{ \underset { x \in \mathcal { X } } { \arg \min V ( x ) } \right \} + \gamma.$$
Analyzing state evolution using Lemma 6.1. Let us define the function
$$\ m m s e ( s ) = \mathbb { E } \left \lceil \left ( \bar { \beta } - \mathbb { E } [ \bar { \beta } \, | \, \sqrt { s } \, \bar { \beta } + G ] \right ) ^ { 2 } \right \rceil,$$
where G ∼ N (0 1) , is independent of ¯ β . Then, recalling the definition of f k from (19), the state evolution recursion in (21) is:
$$( \chi _ { c } ^ { k + 1 } ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \underbrace { \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r c } \underbrace { \ m m s e ( ( \chi _ { c } ^ { k } ) ^ { 2 } ) } _ { = \colon \phi _ { r } ^ { k } } } _ { = \colon \phi _ { r } ^ { k } } ) ^ { - 1 }. \\ \cdot \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }. \hat { \ }.$$
Using the definitions above, the state evolution recursion can be rewritten as:
$$\phi _ { r } ^ { k } = \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \sum _ { c = 1 } ^ { C } \widetilde { W } _ { r \epsilon } \psi _ { c } ^ { k }, \quad \psi _ { c } ^ { k + 1 } = \text{mmse} \left ( \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r \epsilon } ( \phi _ { r } ^ { k } ) ^ { - 1 } \right ),$$
which combined into one equation gives:
$$\phi _ { r } ^ { k + 1 } = \sigma ^ { 2 } + \frac { 1 } { \delta _ { \text{in} } } \underbrace { \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r } \text{mmse} \left ( \sum _ { r ^ { \prime } = 1 } ^ { R } \widetilde { W } _ { r ^ { \prime } \text{c} } \left ( \phi _ { r ^ { \prime } } ^ { k } \right ) ^ { - 1 } \right ) } _ { = \colon x _ { r } ^ { k + 1 } }.$$
Rewriting the recursion in terms of x k +1 r defined above, we get:
$$x _ { r } ^ { k + 1 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r \epsilon } \text{mmse} \left ( \sum _ { r ^ { \prime } = 1 } ^ { R } \widetilde { W } _ { r ^ { \prime } \epsilon } \left ( \sigma ^ { 2 } + \frac { x _ { r } ^ { k } } { \delta _ { \text{in} } } \right ) ^ { - 1 } \right ), \ \ r \in [ R ].$$
The modified recursion in (70) is an instance of the coupled recursion in (65), which can be seen by taking A = ˜ W ⊤ and
$$f ( y ) = \text{mmse} \left ( \frac { 1 } { \sigma ^ { 2 } } - y \right ), \ \ g ( x ) = \frac { 1 } { \sigma ^ { 2 } } - \frac { 1 } { \sigma ^ { 2 } + x / \delta }.$$
It is shown in [37, Section VI.E] that with these functions, which satisfy the assumptions stated at the start of this section, the potential function V ( x ) in (66) equals U b δ ( ; in ) defined in (31) (upto an additive constant). Invoking Lemma 6.1, we have that the fixed points of (70), denoted by ( x ∞ r ) r ∈ [ R ] satisfy:
$$\max _ { r \in [ R ] } x _ { r } ^ { \infty } \leq \max \left \{ \underset { x \in [ 0, \text{mmse} ( 0 ) ] } { \arg \min } V ( x ) \right \} + \gamma = \max \left \{ \underset { b \in [ 0, V \text{ar} ( \bar { \beta } ) ] } { \arg \min } U ( b ; \delta _ { \text{in} } ) \right \} + \gamma,$$
where the last equality uses the fact that mmse (0) = Var[ ¯ ] β .
We now use the bound on x ∞ r to upper bound the asymptotic MSE. Using (25), the asymptotic MSE (as k → ∞ ) can be written as 1 C ∑ C c =1 mmse ( ( χ ∞ c ) 2 ) which can be further written as 1 C ∑ C c =1 ψ ∞ c using (69). From (69) and (70), we can write x ∞ r = ∑ C c =1 ˜ W ψ rc ∞ c , which can be written more explicitly as
$$\nmid \geq c = 1 + c \, \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { m } { \dots } { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } { \stackrel { \dots } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } ) } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } { } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } \ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } _ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } \\ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } }. } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } & } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } | } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } ( } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } }, } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } [ } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } ; } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } } = \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \$$
For notational convenience, let us denote x ∗ = max { argmin b ∈ [0 Var( ¯ ) , β U b δ ( ; in ) } + γ . From (71), we have that x ∞ r ≤ x ∗ for all r ∈ [ R ] where R = Λ + ω -1 . In the rightmost vector in (72), we observe that each entry contains the sum of at most ω consecutive terms. This implies that
$$\sum _ { c ^ { \prime } = c } ^ { c + \omega - 1 } \psi _ { c ^ { \prime } } ^ { \infty } \leq x ^ { * } \omega, \ \ c \in [ C ].$$
Recalling that C = Λ and dividing the elements of [ ψ ∞ 1 , . . . , ψ ∞ Λ ] into groups of non-intersecting consecutive terms - with index groups [1 : ω , ] [ ω +1 : 2 ω , . . . , ] [Λ -ω +1 : Λ] - gives us at most
⌈ Λ ω ⌉ disjoint groups, with the sum of each group having an upper bound of x ω ∗ by (73). Hence, the asymptotic MSE can be bounded as
$$\frac { 1 } { C } \sum _ { c = 1 } ^ { C } \psi _ { c } ^ { \infty } \leq \frac { 1 } { \Lambda } \left \lceil \frac { \Lambda } { \omega } \right \rceil x ^ { * } \omega < \frac { 1 } { \Lambda } \left ( \frac { \Lambda } { \omega } + 1 \right ) x ^ { * } \omega = \frac { \Lambda + \omega } { \Lambda } x ^ { * },$$
This completes the proof of the first part of Theorem 3.5.
## 6.2 Proof of (33)
For the i.i.d. design, we have R = C = 1 and ˜ W 11 = 1 , so the state evolution reduces to
$$x ^ { k + 1 } = m m s e \left ( \left ( \sigma ^ { 2 } + \frac { x ^ { k } } { \delta } \right ) ^ { - 1 } \right ),$$
with the initialization x 0 = mmse (0) = Var( ¯ ) β . Since mmse ( s ) is strictly decreasing in s ∈ [0 , ∞ ) , the sequence ( x k ) is monotonically decreasing in k , and since it is bounded below, it converges to a fixed point. Since the recursion is initialized at x 0 = Var( ¯ ) β , the fixed point is given by the largest solution of x = mmse ( ( σ 2 + x δ ) -1 ) . Finally, we observe that the same equation is obtained by setting ∂ U b δ 1 ( ; ) = 0 . This completes the proof of (33).
## 7 Discussion and Future Directions
We have shown that for noiseless QGT and pooled data, a spatially coupled Bernoulli test design with an AMP recovery algorithm achieves almost-exact recovery with n = o p ( ) tests. A key open question is to determine how n scales with p for almost-exact recovery with SC-AMP. Deriving this scaling is beyond the reach of our asymptotic analysis, which requires that n/p → δ > 0 , but recent nonasymptotic analyses of AMP [29,60] might provide tools to address this question, and allow us to compare with information-theoretic bound of n ∗ = γ ∗ p log p (see (4)).
̸
In this paper, the only assumption on the QGT signal vector β is that its empirical distribution converges to a Bernoulli distribution. The items are not required to be independent, and in some applications there may be known correlations between the items. Although the current SC-AMP algorithm does not exploit correlations between the signal entries, it can be adapted to do so, using nonseparable denoising functions [61]. Recent work in Boolean group testing has shown that exploiting correlations or community structure among the items can significantly improve testing efficiency [62, 63]. Exploring how community structure can be harnessed in quantitative group testing is an interesting direction for future work.
Another open question is to determine the number of tests required for exact recovery in the linear regime for an efficient scheme with a random design. We recall that exact recovery requires P [ ˜ = β β ] → 0 as p →∞ , in contrast to the almost-exact recovery criterion in (2).
## A Proof of Theorem 5.2
## A.1 Proof of Theorem 5.2 via Reduction to Abstract Matrix-AMP
We describe an abstract matrix-AMP iteration for which a state evolution result can be established, and then prove Theorem 5.2 by reducing the SC-GAMP algorithm to the abstract matrix-AMP. For k ≥ 0 , the abstract matrix-AMP produces iterates H k +1 ∈ R p × l H and E k +1 ∈ R n × l E as follows:
$$H ^ { k + 1 } = \widetilde { X } ^ { \top } \widehat { R } ^ { k } - \widehat { H } ^ { k } \cdot ( { \mathbf D } ^ { k } ) ^ { \top }, \ \widehat { R } ^ { k } = \tilde { g } _ { k } ( E ^ { k }, \gamma, \mathcal { R } ), \ \ \mathbf D ^ { k } = \frac { 1 } { \mathbf R } \sum _ { r = 1 } ^ { \mathbf R } \mathbb { E } [ \tilde { g } ^ { \prime } _ { k } ( \tilde { E } ^ { k } _ { r }, \tilde { \gamma }, r ) ]$$
$$E ^ { k + 1 } = \widetilde { X } \widehat { H } ^ { k + 1 } - \widehat { R } ^ { k } \cdot ( B ^ { k + 1 } ) ^ { T }, \ \widehat { H } ^ { k + 1 } = \tilde { f } _ { k + 1 } ( H ^ { k + 1 }, \beta, C ), \ \ B ^ { k + 1 } = \frac { 1 } { \delta C } \sum _ { c = 1 } ^ { c } \mathbb { E } [ \tilde { f } ^ { \prime } _ { k + 1 } ( \tilde { H } ^ { k + 1 } _ { \epsilon }, \bar { \beta }, c ) ]$$
$$H ^ { k + 1 } & = \widetilde { X } ^ { \top } \widehat { R } ^ { k } - \widehat { H } ^ { k } \cdot ( { \mathbf D } ^ { k } ) ^ { \top }, \quad \widehat { R } ^ { k } = \tilde { g } _ { k } ( E ^ { k }, \gamma, \mathcal { R } ), \quad { \mathbf D } ^ { k } = \frac { 1 } { R } \sum _ { r = 1 } ^ { \bullet } \mathbb { E } [ \tilde { g } _ { k } ^ { \prime } ( \bar { E } _ { r } ^ { k }, \bar { \gamma }, r ) ] \\ E ^ { k + 1 } & = \widetilde { X } \widehat { H } ^ { k + 1 } - \widehat { R } ^ { k } \cdot ( { \mathbf B } ^ { k + 1 } ) ^ { \top }, \quad \widehat { H } ^ { k + 1 } = \tilde { f } _ { k + 1 } ( H ^ { k + 1 }, \beta, \mathcal { C } ), \quad { \mathbf B } ^ { k + 1 } = \frac { 1 } { \delta \mathbb { C } } \sum _ { \varepsilon - 1 } ^ { \mathbb { C } } \mathbb { E } [ \tilde { f } _ { k + 1 } ^ { \prime } ( \bar { H } _ { \varepsilon } ^ { k + 1 }, \bar { \zeta } }$$
where β ∈ R p , γ ∈ R n , and C and R are defined in (51). The functions ˜ f k +1 : R l H × R × [ C ] → R l E and ˜ g k : R l E × R × [ R ] → R l H act row-wise on their inputs. The joint laws of ( ¯ E ,γ k r ¯) and ( ¯ H k +1 c , β ¯ ) are described later (below (75)). The algorithm is initialized with ̂ H 0 ∈ R p × l H and E 0 = ˜ ̂ XH 0 ∈ R n × l E . We have the following assumptions for the abstract matrix-AMP algorithm.
- (C1) As dimensions p, n →∞ , the ratio n/p → δ > 0 . Furthermore, l E , l H , R , and C are positive integers that do not scale with p as n, p →∞ .
- (C2) Almost surely for all c ∈ [ C ] , as n, p → ∞ , ( β , H c ̂ 0 J c , : ) W → ( ¯ β, H ¯ 0 c ) and γ r W → ¯ γ , with the joint law of ( ¯ β, H ¯ 0 c ) ∈ R × R l E having finite moments of all orders, where J c is defined in (17). Multivariate polynomials are dense in the real L 2 -spaces of functions f : R 2 → R and g : R l E +2 → R with the inner products
$$\langle f, \tilde { f } \rangle \coloneqq \mathbb { E } [ f ( \bar { \gamma }, r ) \tilde { f } ( \bar { \gamma }, r ) ] \ \text{ and } \ \langle g, \tilde { g } \rangle \coloneqq \mathbb { E } [ g ( \bar { \beta }, \bar { H } _ { \mathfrak { c } } ^ { 0 }, \mathfrak { c } ) \tilde { g } ( \bar { \beta }, \bar { H } _ { \mathfrak { c } } ^ { 0 }, \mathfrak { c } ) ],$$
for all r ∈ [ R ] and c ∈ [ C ] .
- (C3) For k ≥ 0 , the functions ˜ f k +1 and ˜ g k are continuous, Lipschitz w.r.t. their first argument, and satisfy the polynomial growth condition in (5) for some order r ≥ 1 .
- (C4) ˜ X is a generalized white noise matrix where ∥ ˜ X ∥ op < C almost surely for sufficiently large n, p for some constant C . For any fixed polynomial functions f † : R l E +2 → R and f ‡ : R 2 → R , as n, p →∞ ,
$$\max _ { i \in [ n ] } \left | \left \langle f ^ { \dagger } ( \beta _ { \mathrm c }, \widehat { H } _ { \mathcal { I } _ { \mathrm c } } ^ { 0 }, \mathrm c ) \odot S _ { i, \mathcal { I } _ { \mathrm c } } \right \rangle - \left \langle f ^ { \dagger } ( \beta _ { \mathrm c }, \widehat { H } _ { \mathcal { I } _ { \mathrm c } } ^ { 0 }, \mathrm c ) \right \rangle \cdot \langle S _ { i, \mathcal { I } _ { \mathrm c } } \rangle \right | \stackrel { a. s. } { \to } 0 \\ \max _ { j \in [ p ] } \left | \left \langle f ^ { \dagger } ( \gamma _ { r }, r ) \odot S _ { \mathcal { I } _ { r }, j } \right \rangle - \left \langle f ^ { \dagger } ( \gamma _ { r }, r ) \right \rangle \cdot \langle S _ { \mathcal { I } _ { r }, j } \right \rangle \right | \stackrel { a. s. } { \to } 0,$$
for all c ∈ [ C ] and r ∈ [ R ] , where S is the variance profile of ˜ X (see Definition 5.1).
State evolution. The state evolution parameters for k ≥ 0 are
$$\Omega ^ { k + 1 } & = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \widehat { \Omega } ^ { k + 1, r } ; \quad \widehat { \Omega } ^ { k + 1, r } = \mathbb { E } [ \tilde { g } _ { k } ( \bar { E } _ { r } ^ { k }, \bar { \gamma }, r ) \tilde { g } _ { k } ( \bar { E } _ { r } ^ { k }, \bar { \gamma }, r ) ^ { \top } ] \in \mathbb { R } ^ { l _ { H } \times l _ { H } } \\ \Pi ^ { k + 1 } & = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \widehat { \Pi } ^ { k + 1, c } ; \quad \widehat { \Pi } ^ { k + 1, c } = \frac { 1 } { \delta } \mathbb { E } [ \tilde { f } _ { k + 1 } ( \bar { H } _ { c } ^ { k + 1 }, \bar { \beta }, c ) \tilde { f } _ { k + 1 } ( \bar { H } _ { c } ^ { k + 1 }, \bar { \beta }, c ) ^ { \top } ] \in \mathbb { R } ^ { l _ { E } \times l _ { E } }, \\ -. & \quad. \quad - \dots \quad - \dots \quad \dots \quad -$$
with ¯ E k r ∼ N (0 , Π ) k independent of ¯ γ , and ¯ H k +1 c ∼ N (0 Ω , k +1 ) independent of ¯ β . The state evolution is initialized with
$$\Pi ^ { 0 } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \widehat { \Pi } ^ { 0, c }, \quad \text{where} \quad \widehat { \Pi } ^ { 0, c } = \frac { 1 } { \delta } \mathbb { E } \left [ \bar { H } _ { c } ^ { 0 } ( \bar { H } _ { c } ^ { 0 } ) ^ { \top } \right ].$$
Theorem A.1 (State evolution for abstract matrix-AMP) . Consider the abstract matrix-AMP in (74) with the assumptions (C1)-(C4) being satisfied. For k ≥ 1 , and for r ∈ [ R ] , c ∈ [ C ] , the iterates of the abstract matrix AMP satisfy
$$( H ^ { k } _ { \mathcal { J } _ { c }, \cdot }, \beta _ { c } ) \stackrel { W _ { 2 } } { \to } ( \bar { H } ^ { k } _ { c }, \bar { \beta } ), \ \ ( E ^ { k } _ { \mathcal { I } _ { r }, \cdot }, \gamma _ { r } ) \stackrel { W _ { 2 } } { \to } ( \bar { E } ^ { k } _ { r }, \bar { \gamma } ),$$
where ¯ H k c is independent of ¯ β , and ¯ E k r is independent of ¯ γ .
Theorem A.1 is proved in Section A.2.
Proof of Theorem 5.2 using Theorem A.1. We reduce the SC-GAMP algorithm to the abstract matrix-AMP iteration. As given in (55), we can obtain ˜ X from ˜ X sc , which is a generalized white noise matrix (see Definition 5.1) with variance profile S ij = 1 for all ( i, j ) ∈ [ n ] × [ p ] . Next, we set γ := Ψ ˜ , the same β for both algorithms, and the functions ˜ f k : R C × R × [ C ] → R 2 R and ˜ g k : R 2 R × R × [ R ] → R C as follows:
$$\tilde { f } _ { k } ( H ^ { k } _ { j, \cdot }, \beta _ { j }, c ) = \left \lceil \beta _ { j } \left ( \sqrt { R \widetilde { W } _ { 1 c } }, \dots, \sqrt { R \widetilde { W } _ { R c } } \right ), \ f _ { k } ( H ^ { k } _ { j c } + \mu ^ { k } _ { \beta, c } \beta _ { j }, c ) \left ( \sqrt { R \widetilde { W } _ { 1 c } }, \dots, \sqrt { R \widetilde { W } _ { R c } } \right ) \right \rceil,$$
for j ∈ J c and H k j, : ∈ R C (i.e., l H = C ). We also set
$$\tilde { g } _ { k } ( E _ { i, \cdot } ^ { k }, \gamma _ { i }, r ) = g _ { k } ( E _ { i r } ^ { k }, \, q ( E _ { i, r + R }, \widetilde { \Psi } _ { i } ), \, r ) \left ( \sqrt { R \widetilde { W } _ { r 1 } }, \dots, \sqrt { R \widetilde { W } _ { r C } } \right ),$$
for i ∈ I r and E k i, : ∈ R 2 R (i.e., l E = 2 R ). The abstract matrix-AMP iteration is initialized with
$$\widehat { H } _ { j, \cdot } ^ { 0 } = \left [ \beta _ { j } \left ( \sqrt { R \widetilde { W } _ { 1 c } }, \dots, \sqrt { R \widetilde { W } _ { R c } } \right ), \, \hat { \beta } _ { j } ^ { 0 } \left ( \sqrt { R \widetilde { W } _ { 1 c } }, \dots, \sqrt { R \widetilde { W } _ { R c } } \right ) \right ], \quad \text{for } j \in \mathcal { J } _ { c }.$$
The state evolution parameters Π k ∈ R 2 R × 2 R and Ω k ∈ R C × C are recursively computed as follows. We have ¯ E k r ∼ N (0 , Π ) k independent of ¯ Ψ , and the entries of ̂ Ω k +1 , r are
$$\widehat { \Omega } _ { \text{cc} ^ { \prime } } ^ { k + 1, r } = \text{R} \cdot \mathbb { E } \left [ g _ { k } ( \bar { E } _ { r, r } ^ { k }, q ( \bar { E } _ { r, r + R }, \bar { \Psi } ), r ) ^ { 2 } \right ] \sqrt { \widetilde { W } _ { \text{cc} } \widetilde { W } _ { \text{cc} ^ { \prime } } }, \quad \text{for } c, c ^ { \prime } \in [ C ].$$
Next, we have ¯ H k c ∼ N (0 Ω ) , k independent of ¯ β and
$$\text{iveA, we move } \iota _ { c } \stackrel { \sim } { \sim } \mathcal { I } \left ( \mathcal { I }, \mathcal { I } \right ) \text{unique sequence on } \beta \text{ and } \\ \widehat { \Pi } _ { r s } ^ { k + 1, c } = \begin{cases} \frac { 1 } { \delta } \mathbb { R } \mathbb { E } [ \bar { \beta } ^ { 2 } ] \sqrt { \widetilde { W } _ { r c } \widetilde { W } _ { s c } } & r, s \in [ R ] \\ \frac { 1 } { \delta } \mathbb { R } \mathbb { E } [ \bar { \beta } f _ { k } ( \{ \bar { H } _ { c } ^ { k } \} _ { c } + \mu _ { \beta, c } ^ { k + 1 } \bar { \beta }, c ) ] \sqrt { \widetilde { W } _ { r c } \widetilde { W } _ { ( r - R ) c } } & r \in [ R ], R + 1 \leq r \leq 2 R, s \in [ R ] \\ \frac { 1 } { \delta } \mathbb { R } [ \bar { \beta } f _ { k } ( \{ \bar { H } _ { c } ^ { k } \} _ { c } + \mu _ { \beta, c } ^ { k + 1 } \bar { \beta }, c ) ] \sqrt { \widetilde { W } _ { ( r - R ) c } \widetilde { W } _ { s c } } & R + 1 \leq r \leq 2 R, s \in [ R ] \\ \frac { 1 } { \delta } \mathbb { R } [ f _ { k } ( \{ \bar { H } _ { c } ^ { k } \} _ { c } + \mu _ { \beta, c } ^ { k + 1 } \bar { \beta }, c ) ^ { 2 } ] \sqrt { \widetilde { W } _ { ( r - R ) c } \widetilde { W } _ { ( s - R ) c } } & R + 1 \leq r, s \leq 2 R. \end{cases}$$
The state evolution is initialized with
$$\widehat { \Pi } ^ { 0, \mathfrak { c } } = \frac { 1 } { \delta } \lim _ { p \to \infty } \frac { 1 } { p / \mathbb { C } } ( \widehat { H } ^ { 0 } _ { \mathcal { J } _ { \mathfrak { c } }, \cdot } ) ^ { \top } \widehat { H } ^ { 0 } _ { \mathcal { J } _ { \mathfrak { c } }, \cdot },$$
for c ∈ [ C ] , with ̂ H 0 J c , : ∈ R p/ C × 2 R . By assumption (A2) (see (56)), the entries of ̂ Π 0 , c are given by
$$\widehat { \Pi } _ { r s } ^ { 0, c } & = \begin{cases} \frac { 1 } { \delta } R \mathbb { E } [ \bar { \beta } ^ { 2 } ] \sqrt { \widetilde { W } _ { r c } \widetilde { W } _ { s c } } & r, s \in [ R ] \\ \frac { 1 } { \delta } \mathbb { E } [ \bar { \beta } \bar { \beta } _ { c } ^ { 0 } ] \sqrt { \widetilde { W } _ { r c } \widetilde { W } _ { ( s - R ) c } } & r \in [ R ], R + 1 \leq s \leq 2 R \\ \frac { 1 } { \delta } \mathbb { E } [ \bar { \beta } \bar { \beta } _ { c } ^ { 0 } ] \sqrt { \widetilde { W } _ { ( r - R ) c } \widetilde { W } _ { s c } } & R + 1 \leq s \leq 2 R, s \in [ R ] \\ \frac { 1 } { \delta } \mathbb { E } [ ( \bar { \beta } _ { c } ^ { 0 } ) ^ { 2 } ] \sqrt { \widetilde { W } _ { ( r - R ) c } \widetilde { W } _ { ( s - R ) c } } & R + 1 \leq s, r \leq 2 R. \end{cases} \\ \int \int \int \frac { 1 } { \delta } \equiv c \quad \widehat { \frac { \widehat { \Pi } } { \delta } } c$$
We then have Π = 0 1 ∑ C ̂ Π 0 , c .
C c =1 We can then show that
$$\begin{bmatrix} \Pi _ { r r } ^ { k } & \Pi _ { r ( r + R ) } ^ { k } \\ \Pi _ { ( r + R ) r } ^ { k } & \Pi _ { ( r + R ) ( r + R ) } \end{bmatrix} = \Sigma ^ { k, r } \ \text{ and } \ \Omega _ { \text{cc} } ^ { k + 1 } = ( \sigma _ { \beta, \text{c} } ^ { k + 1 } ) ^ { 2 },$$
implying that ( { ¯ E k r } r , { ¯ E k r } r + R ) d = ( Z , Z r k r ) and { ¯ H k c } c d = G k β, c . We can also show that for k ≥ 0 ,
$$& E _ { i, r } ^ { k } = \Theta _ { i } ^ { k }, \quad E _ { i, r + R } ^ { k } = \Theta _ { i }, \quad \text{for $i\in\mathcal{I}_{r}$, $r\in[R]$} \\ & H _ { j, c } ^ { k + 1 } + \mu _ { \beta, c } ^ { k + 1 } \beta _ { j } = \beta _ { j } ^ { k + 1 }, \quad \text{for $j\in\mathcal{J}_{c}$, $c\in[C]$.}$$
Both (77) and (78) are shown using steps identical to those in [35, Section 5.1.2], so we omit repeating the proof for brevity. Theorem 5.2 follows by using (77) and (78) in Theorem A.1.
## A.2 Proof of Theorem A.1 via Reduction to U-AMP
Theorem A.1 is proved by reducing the abstract matrix-AMP recursion to the U-AMP recursion which is defined as follows. Given a generalized white noise matrix ˜ X , for t ≥ 1 , the iterates of U-AMP, denoted by h t ∈ R p and e t ∈ R n , are produced using functions f v t : R t + L d +1 → R , f u t +1 : R t + L c +1 → R . Given an initializer u 1 ∈ R n , side information vectors c 1 , . . . , c L c ∈ R n and d , . . . , d 1 L d ∈ R p , all independent of ˜ X , the iterates of the U-AMP recursion are computed as:
$$, \dots, \infty \, & \ < a \,, \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \min \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \max \, \text{$\mathcal{ } } \\ h ^ { t } & = \sqrt { \delta } \widetilde { X } ^ { \top } u ^ { t } - \sum _ { s = 1 } ^ { t } b ^ { t } _ { s } v ^ { s }, \quad \ \ \ \ \ v ^ { t } = f ^ { v } _ { t + 1 } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { L _ { d } }, \mathcal { C } ) \\ e ^ { t } & = \sqrt { \delta } \widetilde { X } v ^ { t } - \sum _ { s = 1 } ^ { t } a ^ { t } _ { s } u ^ { s }, \quad \ \ \ u ^ { t + 1 } = f ^ { u } _ { t + 1 } ( e ^ { 1 }, \dots, e ^ { t }, c ^ { 1 }, \dots, e ^ { L _ { c } }, \mathcal { R } ),$$
where C R , were defined in (51), and the functions f v t and f u t +1 act row-wise. The coefficients a t s and b t s are defined later in (83) in terms of state evolution parameters.
Recalling the notification simplification for sub-blocks of vectors presented in the paragraph below (17), we have the following assumptions:
- (D1) As n, p → ∞ , we have n/p = δ > 0 , for fixed L c and L d . Furthermore, for all c ∈ [ C ] and r ∈ [ R ] ,
$$( u _ { r } ^ { 1 }, c _ { r } ^ { 1 }, \dots, c _ { r } ^ { L _ { c } } ) \stackrel { W } { \to } ( \bar { u } _ { r } ^ { 1 }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } } ) \text{ and } ( d _ { c } ^ { 1 }, \dots, d _ { c } ^ { L _ { d } } ) \stackrel { W } { \to } ( \bar { d } _ { c } ^ { 1 }, \dots, \bar { d } _ { c } ^ { L _ { d } } ),$$
for joint limit laws (¯ u , c 1 r ¯ 1 r , . . . , c ¯ L c r ) and ( ¯ d , . . . , d 1 c ¯ L d c ) having finite moments of all orders, where E [(¯ u 1 ) 2 ] ≥ 0 . Multivariate polynomials are dense in the real L 2 -spaces of functions f : R L c +1 → R and g : R L d → R with the inner products
$$\langle f, \tilde { f } \rangle \coloneqq \mathbb { E } [ f ( \bar { u } _ { r } ^ { 1 }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } } ) \tilde { f } ( \bar { u } _ { r } ^ { 1 }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } } ) ] \text{ and } \langle g, \tilde { g } \rangle \coloneqq \mathbb { E } [ g ( \bar { d } _ { c } ^ { 1 }, \dots, \bar { d } _ { c } ^ { L _ { d } } ) \tilde { g } ( \bar { d } _ { c } ^ { 1 }, \dots, \bar { d } _ { c } ^ { L _ { d } } ) ].$$
- (D2) Each function f v t : R t + L d +1 → R and f u t +1 : R t + L c +1 → R is continuous, is Lipschitz in its first t arguments, and satisfies the polynomial growth condition in (5) for some order r ≥ 1 .
- (D3) ∥ ˜ X ∥ op < C , for some constant C almost surely for all sufficiently large n and p .
- (D4) For any fixed polynomial functions f † : R L c +1 → R and f ‡ : R L d → R , almost surely as n, p →∞ ,
$$\max _ { j \in [ p ] } & \left | \langle f ^ { \dagger } ( u ^ { 1 }, c ^ { 1 }, \dots, c ^ { L _ { c } } ) \odot S _ { ;, j } \rangle - \langle f ^ { \dagger } ( u ^ { 1 }, c ^ { 1 }, \dots, c ^ { L _ { c } } ) \rangle \cdot \langle S _ { ;, j } \rangle \right | \rightarrow 0 \\ & \max _ { i \in [ n ] } \left | \langle f ^ { \dagger } ( d ^ { 1 }, \dots, d ^ { L _ { d } } ) \odot S _ { i, ; } \rangle - \langle f ^ { \dagger } ( d ^ { 1 }, \dots, d ^ { L _ { d } } ) \rangle \cdot \langle S _ { i, ; } \rangle \right | \rightarrow 0.$$
State evolution. The state evolution result below states that the joint empirical distribution of ( h , . . . , h 1 c t c ) converges to a Gaussian law N (0 Ξ ) , t , for c ∈ [ C ] . Similarly, the joint empirical distribution of ( e , . . . , e 1 r t r ) converges to N (0 , Γ ) t , for r ∈ [ R ] . The covariance matrices Ξ t , Γ t ∈ R t × t are iteratively defined as follows, starting from Ξ 1 = δ R ∑ R r =1 E [(¯ u 1 r ) 2 ] . Given Ξ t , for t ≥ 1 , let ( ¯ h , . . . , h 1 ¯ ) t ∼ N (0 , Ξ ) t be independent of ( ¯ d , . . . , d 1 c ¯ L d c ) and define
$$\bar { v } _ { \text{c} } ^ { s } = f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { s }, \bar { d } _ { \text{c} } ^ { 1 }, \dots, \bar { d } _ { \text{c} } ^ { L _ { d } }, \text{c} ), \quad s \in [ t ], \text{c} \in [ \text{C} ].$$
Then we have
$$\Gamma ^ { t } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \widehat { \Gamma } ^ { t, c } \quad \text{where} \quad \widehat { \Gamma } ^ { t, c } = \left ( \mathbb { E } [ \bar { v } _ { c } ^ { r } \bar { v } _ { c } ^ { s } ] \right ) _ { r, s = 1 } ^ { t }.$$
Next, let (¯ e , . . . , e 1 ¯ ) t ∼ N (0 , Γ ) t be independent of (¯ u , c 1 r ¯ 1 r , . . . , c ¯ L c r ) and define
$$\bar { u } _ { r } ^ { s + 1 } = f _ { s + 1 } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { s }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } }, r ), \ \ s \in [ t ], \, r \in [ \mathbf R ].$$
Then, we have
$$\Xi ^ { t + 1 } = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \widehat { \Xi } ^ { t + 1, r } \ \text{ where } \ \widehat { \Xi } ^ { t + 1, r } = \delta \left ( \mathbb { E } [ \bar { u } ^ { r } _ { r } \bar { u } ^ { s } _ { r } ] \right ) ^ { t + 1 } _ { r, s = 1 }.$$
We define the memory coefficients a t s and b t s in (79) as:
$$a _ { s } ^ { t } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \mathbb { E } \left [ \partial _ { s } f _ { t } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } _ { c } ^ { 1 } \dots, \bar { d } _ { c } ^ { L _ { d } }, c ) \right ], \quad b _ { s } ^ { t } = \frac { \delta } { R } \sum _ { r = 1 } ^ { R } \mathbb { E } \left [ \partial _ { s } f _ { t } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } }, r ) \right ],$$
where ∂ s denotes partial derivative in the s th argument. The following result gives the state evolution result for the U-AMP recursion.
Corollary A.2 (State evolution for U-AMP) . Let ˜ X ∈ R n × p be a generalized white noise matrix (as defined in Definition 5.1) with variance profile S ∈ R n × p , and let u , c 1 r 1 r , . . . , c L c r , d 1 c , . . . , d L d c be independent of ˜ X and satisfy Assumptions (D1)-(D4). Further assume that each matrix Ξ t and Γ t is non-singular. Then for any fixed t ≥ 1 , as n, p → ∞ , the iterates of the abstract AMP in (79) almost surely satisfy the following, for r ∈ [ R ] and c ∈ [ C ] :
$$& ( u _ { r } ^ { 1 }, c _ { r } ^ { 1 }, \dots, c _ { r } ^ { L _ { c } }, e _ { r } ^ { 1 }, \dots, e _ { r } ^ { t } ) \stackrel { W _ { 2 } } { \to } ( \bar { u } _ { r } ^ { 1 }, \bar { c } _ { r } ^ { 1 }, \dots, \bar { c } _ { r } ^ { L _ { c } }, \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t } ) \\ & ( d _ { \mathbf c } ^ { 1 }, \dots, d _ { \mathbf c } ^ { L _ { d } }, h _ { \mathbf c } ^ { 1 }, \dots, h _ { \mathbf c } ^ { t } ) \stackrel { W _ { 2 } } { \to } ( \bar { d } _ { \mathbf c } ^ { 1 }, \dots, \bar { d } _ { \mathbf c } ^ { L _ { d } }, \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t } ),$$
where ( ¯ h , . . . , h 1 ¯ ) t ∼ N (0 , Ξ ) t and (¯ e , . . . , e 1 ¯ ) t ∼ N (0 , Γ ) t are independent of ( ¯ d , . . . , d 1 c ¯ L d c ) and (¯ u , c 1 r ¯ 1 r . . . . , c ¯ L c r ) respectively.
Corollary A.2 is obtained from the AMP universality result in [36, Theorem 2.17]. A proof is provided in Section A.3.
Proof of Theorem A.1 using Corollary A.2. We reduce the abstract matrix-AMP iteration to the U-AMP. We set the initializer to be u 1 = 0 , L c = 1 , L d = l E +1 and the side information vectors to be
$$c ^ { 1 } = \gamma, \ \ d ^ { 1 } = \beta, \ \ d ^ { 2 } = \widehat { H } ^ { 0 } _ { \cdot ; 1 }, \ \ \dots, \ \ d ^ { l _ { E } + 1 } = \widehat { H } ^ { 0 } _ { \cdot ; l _ { E } }.$$
We show the reduction through induction.
Base case. We consider the case t = 0 , and our goal is to reduce ̂ H 0 , E 0 , H 1 , ̂ R 0 , ̂ H 1 , and E 1 to iterates of U-AMP defined via careful choices of the functions f v t and f u t +1 . We provide a summary of the reductions before giving their derivations.
- · For t = 1 , . . . , l E : We have
$$h ^ { 1 }, \dots, h ^ { l _ { E } } = 0, \ \ ( v ^ { 1 }, \dots, v ^ { l _ { E } } ) = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 0 }, \ \ ( e ^ { 1 }, \dots, e ^ { l _ { E } } ) = E ^ { 0 }, \ \ u ^ { 1 }, \dots, u ^ { l _ { E } } = 0. \quad \ ( 8 5 )$$
- · For t = l E +1 , . . . , l E + l H : We have
$$& ( h ^ { l _ { E } + 1 }, \dots, h ^ { l _ { E } + l _ { H } } ) = H ^ { 1 }, \quad v ^ { l _ { E } + 1 }, \dots, v ^ { l _ { E } + l _ { H } } = 0 \\ & e ^ { l _ { E } + 1 }, \dots, e ^ { l _ { E } + l _ { H } } = 0, \quad ( u ^ { l _ { E } + 1 }, \dots, u ^ { l _ { E } + l _ { H } } ) = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { 0 }.$$
- · For t = l E + l H +1 , . . . , 2 l E + l H : We have
$$h ^ { l _ { E } + l _ { H } + 1 }, \dots, h ^ { 2 l _ { E } + l _ { H } } = 0, \quad ( v ^ { l _ { E } + l _ { H } + 1 }, \dots, v ^ { 2 l _ { E } + l _ { H } } ) = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 1 } \\ ( e ^ { l _ { E } + l _ { H } + 1 }, \dots, e ^ { 2 l _ { E } + l _ { H } } ) = E ^ { 1 }, \quad u ^ { l _ { E } + l _ { H } + 1 }, \dots, u ^ { 2 l _ { E } + l _ { H } } = 0.$$
We now provide the derivations of (85)-(87). For t = 1 , . . . , l E -1 , we set
$$f _ { t } ^ { v } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { l _ { E } + 1 }, \mathcal { C } ) = \frac { 1 } { \sqrt { \delta } } d ^ { t + 1 }, \ \ f _ { t + 1 } ^ { u } ( e ^ { 1 }, \dots, e ^ { t }, c ^ { 1 }, \mathcal { R } ) = 0.$$
For t = 1 , we have our initializer u 1 = 0 , and using (84) gives
$$h ^ { 1 } = 0, \ \ v ^ { 1 } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 0 } _ { ; ; 1 }, \ \ e ^ { 1 } = E ^ { 0 } _ { ; ; 1 }, \ \ u ^ { 2 } = 0.$$
Following similar steps, for t = 2 , . . . , l E -1 , we have
$$h ^ { t } = 0, \ \ v ^ { t } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 0 } _ { \colon, t }, \ \ e ^ { t } = E ^ { 0 } _ { \colon, t }, \ \ u ^ { t + 1 } = 0.$$
For t = l E , set
$$f _ { l _ { E } } ^ { v } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { l _ { E } + 1 }, \mathcal { C } ) = \frac { 1 } { \sqrt { \delta } } d ^ { l _ { E } + 1 }, \quad f _ { l _ { E } + 1 } ^ { u } ( \underbrace { e ^ { 1 }, \dots, e ^ { l _ { E } } } _ { = E ^ { 0 } }, \underbrace { e ^ { 1 }, \mathcal { R } } _ { = \gamma } = \frac { 1 } { \sqrt { \delta } } \{ \tilde { g } _ { 0 } ( E ^ { 0 }, \gamma, \mathcal { R } ) \} _ { ; 1 },$$
which gives
$$h ^ { l _ { E } } = 0, \ \ v ^ { l _ { E } } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 0 } _ { \colon, l _ { E } }, \ \ e ^ { l _ { E } } = E ^ { 0 } _ { \colon, l _ { E } }, \ \ u ^ { l _ { E } + 1 } = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { 0 } _ { \colon, 1 }.$$
This completes the derivation for (85). For t = l E +1 , . . . , l E + l H -1 , we set
$$f _ { t } ^ { v } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { l _ { E } + 1 }, \mathcal { C } ) = 0, \ \ f _ { t + 1 } ^ { u } ( e ^ { 1 }, \dots, e ^ { t }, c ^ { 1 }, \mathcal { R } ) = \frac { 1 } { \sqrt { \delta } } \{ \tilde { g } _ { 0 } ( E ^ { 0 }, \gamma, \mathcal { R } ) \} _ { ; t + 1 - l _ { E } }.$$
For t = l E +1 , we have the following identity:
$$\text{for } \iota = \iota _ { E } + 1, \, \text{we have one following} \text{tenuity} \colon \\ \{ \widehat { H } ^ { 0 } ( \mathbb { D } ^ { 0 } ) ^ { \top } \} _ { \vdots ; 1 } & = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \left \{ \widehat { H } ^ { 0 } \left ( \mathbb { E } [ \tilde { g } _ { 0 } ^ { \prime } ( \bar { E } _ { r } ^ { 0 }, \bar { \gamma }, r ) ] \right ) ^ { \top } \right \} _ { \vdots, 1 } = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \widehat { H } ^ { 0 } \cdot \begin{bmatrix} \mathbb { E } [ \partial _ { 1 } \tilde { g } _ { 0, 1 } ( \bar { E } _ { r } ^ { 0 }, \bar { \gamma }, r ) ] \\ \vdots \\ \mathbb { E } [ \partial _ { l } \tilde { g } _ { 0, 1 } ( \bar { E } _ { r } ^ { 0 }, \bar { \gamma }, r ) ] \end{bmatrix} \\ & = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \sum _ { s = 1 } ^ { l _ { E } } \widehat { H } ^ { 0 } _ { \vdots, s } \mathbb { E } [ \partial _ { s } \tilde { g } _ { 0, 1 } ( \bar { E } _ { r } ^ { 0 }, \bar { \gamma }, r ) ].$$
Then, we have
$$h ^ { l _ { E } + 1 } & \overset { ( a ) } { = } \sqrt { \delta } \widetilde { X } ^ { \top } u ^ { l _ { E } + 1 } - \sum _ { s = 1 } ^ { l _ { E } } b ^ { l _ { E } + 1 } _ { s } v ^ { s } \overset { ( b ) } { = } \widetilde { X } ^ { \top } \widehat { R } ^ { 0 } _ { \cdot, 1 } - \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \sum _ { s = 1 } ^ { l _ { E } } \mathbb { E } [ \partial _ { s } \tilde { g } _ { 0, 1 } ( \bar { R } ^ { 0 } _ { r }, \bar { \gamma }, r ) ] \widehat { H } ^ { 0 } _ { \cdot, s } \\ & \overset { ( c ) } { = } \widetilde { X } ^ { \top } \widehat { R } ^ { 0 } _ { \cdot, 1 } - \{ \widehat { H } ^ { 0 } ( \mathbb { D } ^ { 0 } ) ^ { \top } \} _ { \cdot, 1 } = H ^ { 1 } _ { \cdot, 1 },$$
where we use (79) in (a), substitute the definitions of u l E +1 , b l E +1 s , and v s in (b), and apply (90) in (c). Next, we have
$$v ^ { l _ { E } + 1 } = 0, \ \ e ^ { l _ { E } + 1 } = 0, \ \ u ^ { l _ { E } + 2 } = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { 0 } _ { \cdot, 2 }.$$
Similarly for t = l E +2 , . . . , l E + l H -1 , we have
$$h ^ { t } = H ^ { 1 } _ { \colon, t - l _ { E } }, \ \ v ^ { t } = 0, \ \ e ^ { t } = 0, \ \ u ^ { t + 1 } = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { 0 } _ { \colon, t + 1 - l _ { E } }.$$
For t = l E + l H , we set f v t = 0 and f u t +1 = 0 , so that
$$h ^ { l _ { E } + l _ { H } } = H ^ { 1 } _ { ;, l _ { H } }, \ \ v ^ { l _ { E } + l _ { H } } = 0, \ \ e ^ { l _ { E } + l _ { H } } = 0, \ \ u ^ { l _ { E } + l _ { H } + 1 } = 0.$$
This completes the derivation for (86).
For t = l E + l H +1 , . . . , 2 l E + l H -1 , we set
$$f _ { t } ^ { v } ( h ^ { 1 }, \dots, h ^ { l _ { E } }, \underbrace { h ^ { l _ { E } + 1 }, \dots, h ^ { l _ { E } + l _ { H } } } _ { = H ^ { 1 } }, &, h ^ { l _ { E } + l _ { H } + 1 }, \dots, h ^ { t }, \underbrace { d ^ { 1 } } _ { = \beta }, d ^ { 2 }, \dots, d ^ { l _ { E } + 1 }, \mathcal { C } ) = \frac { 1 } { \sqrt { \delta } } \{ \tilde { f } _ { 1 } ( H ^ { 1 }, \beta, \mathcal { C } ) \} _ { ; 1 } \\ \tilde { \psi } _ { 1 } ( 1 + 1 ) \, \gg \, \quad \, \, \,$$
$$f _ { t + 1 } ^ { u } ( e ^ { 1 }, \dots, e ^ { t }, c ^ { 1 }, \mathcal { R } ) = 0.$$
For t = l E + l H +1 , we have
$$h ^ { l _ { E } + l _ { H } + 1 } = 0, \ \ v ^ { l _ { E } + l _ { H } + 1 } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 1 } _ { \colon, 1 }.$$
We pause to show an identity:
$$\tt v c { \, \rho a v e c \, \tt v a n w a n w a n w a n w a n }. \\ \{ \widehat { R } ^ { 0 } ( B ^ { 1 } ) ^ { T } \} _ { \, \cdots, 1 } & = \frac { 1 } { \delta C } \sum _ { c = 1 } ^ { C } \left \{ \widehat { R } ^ { 0 } \mathbb { E } [ \tilde { f } ^ { t } _ { 1 } ( \tilde { H } ^ { 1 } _ { c }, \tilde { \beta }, c ) ] ^ { T } \right \} _ { \, \cdots, 1 } = \frac { 1 } { \delta C } \sum _ { c = 1 } ^ { C } \widehat { R } ^ { 0 } \begin{bmatrix} \mathbb { E } [ \partial _ { 1 } \tilde { f } _ { 1, 1 } ( \tilde { H } ^ { 1 } _ { c }, \tilde { \beta }, c ) ] \\ \vdots \\ \mathbb { E } [ \partial _ { 1 } \tilde { f } _ { 1, 1 } ( \tilde { H } ^ { 1 } _ { c }, \tilde { \beta }, c ) ] \end{bmatrix} \\ \tt T h o n \, \ f r o n \, \ w a n h a n$$
Then, from (79), we have
$$e ^ { l _ { E } + l _ { H } + 1 } & = \sqrt { \delta } \widetilde { X } v ^ { l _ { E } + l _ { H } + 1 } - \sum _ { s = 1 } ^ { l _ { E } + l _ { H } + 1 } a _ { s } ^ { l _ { E } + l _ { H } + 1 } u ^ { s } \stackrel { ( a ) } { = } \widetilde { X } \widehat { H } ^ { 1 } _ { ; ; 1 } - \frac { 1 } { \delta C } \sum _ { c = 1 } ^ { C } \sum _ { s = 1 } ^ { l _ { H } } \widehat { R } ^ { 0 } _ { ; ; s } \mathbb { E } [ \partial _ { s } \widetilde { f } _ { 1, 1 } ( \widehat { H } ^ { 1 } _ { c ; ; }, \widehat { \beta }, c ) ] \\ & \stackrel { ( b ) } { = } \widetilde { X } \widehat { H } ^ { 1 } _ { ; ; 1 } - \left \{ \widehat { R } ^ { 0 } ( B ^ { 1 } ) ^ { \top } \right \} _ { ; ; 1 } = E ^ { 1 } _ { ; ; 1 },$$
where (a) uses the definitions of v l E + l H +1 , a l E + l H +1 s and u s , and (b) uses (92). Next, we have u l E + l H +1 = 0 . Similarly, for t = l E + l H +2 , . . . , 2 l E + l H -1 , we have
$$h ^ { t } = 0, \ \ v ^ { t } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 1 } _ { \colon t - l _ { E } - l _ { H } }, \ \ e ^ { t } = E ^ { 1 } _ { \colon ; t - l _ { E } - l _ { H } }, \ \ u ^ { t + 1 } = 0.$$
For t = 2 l E + l H , we set
$$f _ { 2 l _ { E } + l _ { H } } ^ { v } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { l _ { E } + 1 }, \mathcal { C } ) = \frac { 1 } { \sqrt { \delta } } \{ \tilde { f } _ { 1 } ( H ^ { 1 }, \beta, \mathcal { C } ) \} _ { \colon l _ { E } }$$
$$f _ { 2 l _ { E } + l _ { H } + 1 } ^ { u } ( e ^ { 1 }, \dots, e ^ { l _ { E } + l _ { H } }, \underbrace { e ^ { l _ { E } + l _ { H } + 1 }, \dots, e ^ { 2 l _ { E } + l _ { H } } } _ { = E ^ { 1 } }, \underbrace { c ^ { 1 } _ { \Lambda } } _ { = \gamma }, \mathcal { R } ) = \frac { 1 } { \sqrt { \delta } } \left \{ \tilde { g } _ { 1 } ( E ^ { 1 }, \gamma, \mathcal { R } ) \right \} _ { ; 1 }.$$
$$& d ^ { * E ^ { \top \ast } }, \mathcal { C } ) = \frac { } { \sqrt { \delta } } \{ f _ { 1 } ( H ^ { * }, \beta, \mathcal { C } ) \} _ { \colon, l _ { E } } \\ & \frac { _ { L E ^ { + } l _ { H } + 1 } } { _ { = E ^ { 1 } } }, \underbrace { c ^ { 1 } } _ { = \gamma }, \mathcal { R } ) = \frac { 1 } { \sqrt { \delta } } \{ \tilde { g } _ { 1 } ( E ^ { 1 }, \gamma, \mathcal { R } ) \} _ { \colon, 1 }.$$
This gives
$$h ^ { 2 l _ { E } + l _ { H } } = 0, \ \ v ^ { 2 l _ { E } + l _ { H } } = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { 1 } _ { \colon, l _ { E } }, \ \ e ^ { 2 l _ { E } + l _ { H } } = E ^ { 1 } _ { \colon, l _ { E } }, \ \ u ^ { 2 l _ { E } + l _ { H } + 1 } = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { 1 } _ { \colon, 1 },$$
completing the derivation of (87). This concludes the reduction of the abstract matrix-AMP iterates to the U-AMP iterates for the case of t = 0 .
We now show the convergence statements in Theorem A.1 for E , H 0 1 , and E 1 by reducing the abstract matrix-AMP SE parameters to the corresponding U-AMP SE parameters.
Convergence of ( E , γ 0 ) . From Assumption (D1) and Corollary A.2, we have ( d , . . . , d 1 c L d c , h , . . . , h 1 c l E c ) W → ( ¯ d , . . . , d 1 c ¯ L d c , h ¯ 1 , . . . , h ¯ l E ) , for c ∈ [ C ] . Recalling that
$$v _ { \mathfrak { c } } ^ { s } = f _ { s } ^ { v } ( h ^ { 1 }, \dots, h ^ { s }, d _ { \mathfrak { c } } ^ { 1 }, \dots, d _ { \mathfrak { c } } ^ { L _ { d } }, \mathfrak { c } ), \quad \bar { v } _ { \mathfrak { c } } ^ { s } = f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { s }, \bar { d } _ { \mathfrak { c } } ^ { 1 }, \dots, \bar { d } _ { \mathfrak { c } } ^ { L _ { d } }, \mathfrak { c } ), \quad \text{for } s \geq 1,$$
the convergence above implies that ( v , . . . , v 1 c l E c ) W → (¯ v , . . . , v 1 c ¯ l E c ) . Since we have shown in (85) that ( v , . . . , v 1 l E ) = 1 √ δ ̂ H 0 , we must have (¯ v , . . . , v 1 c ¯ l E c ) = 1 √ δ ¯ H 0 c , where the latter is given by Assumption (C2).
For r ∈ [ R ] , Corollary A.2 implies that ( e , . . . , e 1 r l E r ) W 2 → (¯ e , . . . , e 1 ¯ l E ) ∼ N (0 Γ , l E ) , where
$$\Gamma ^ { l _ { E } } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \widehat { \Gamma } ^ { l _ { E }, c }, \quad \widehat { \Gamma } ^ { l _ { E }, c } = \left ( \mathbb { E } [ \bar { v } _ { c } ^ { r } \bar { v } _ { c } ^ { s } ] \right ) _ { r, s = 1 } ^ { l _ { E } }.$$
Using (¯ v , . . . , v 1 c ¯ l E c ) = 1 √ δ ¯ H 0 c in (76), we have that ̂ Π 0 , c = Γ ̂ l E , c and Π = Γ 0 l E . Since ( e , . . . , e 1 r l E r ) = E 0 I r , : from (85), and Corollary A.2 guarantees that ( e , . . . , e 1 r l E r ) W 2 → (¯ e , . . . , e 1 ¯ l E ) ∼ N (0 , Γ l E ) , we have E 0 I r , : W 2 → ¯ E 0 r ∼ N (0 , Π ) 0 . Since c 1 = γ , and γ is independent of E 0 , we can use Assumption (C2) and apply Corollary A.2 to ( E 0 I r , : , γ r ) to obtain
$$( E _ { \mathcal { I } _ { r } ; } ^ { 0 }, \gamma _ { r } ) \stackrel { W _ { 2 } } { \to } ( \bar { E } _ { r } ^ { 0 }, \bar { \gamma } ).$$
Convergence of ( H ,β 1 ) . For r ∈ [ R ] , recall that u s +1 r = f u s +1 ( e , . . . , e 1 r s r , c 1 r , . . . , c L c r , r ) and ¯ u s +1 r = f u s +1 (¯ e , . . . , e 1 ¯ s , c ¯ 1 r , . . . , c ¯ L c r , r ) . Corollary A.2 implies ( u , . . . , u 1 r l E + l H r ) W 2 → (¯ u , . . . , u 1 r ¯ l E + l H r ) . Moreover, we have shown in (86) that ( u l E +1 , . . . , u l E + l H ) = 1 √ δ ̂ R 0 = 1 √ δ ˜ ( g 0 E , γ, 0 R ) . Hence, using (95) and noting that ˜ g 0 satisfies the polynomial growth condition in (5), we have (¯ u l E +1 r , . . . , u ¯ l E + l H r ) = 1 √ δ ˜ ( ¯ g 0 E , γ, 0 r ¯ r ) , for r ∈ [ R ] .
Corollary A.2 states that ( h , . . . , h 1 c 2 l E + l H c ) W 2 → ( ¯ h , . . . , h 1 ¯ 2 l E + l H ) ∼ N (0 Ξ , 2 l E + l H ) , where
$$\Xi ^ { 2 l _ { E } + l _ { H } } = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \widehat { \Xi } ^ { 2 l _ { E } + l _ { H }, r }, \quad \widehat { \Xi } ^ { 2 l _ { E } + l _ { H } } = \delta \left ( \mathbb { E } [ \bar { u } ^ { r } _ { r } \bar { u } ^ { s } _ { r } ] \right ) ^ { 2 l _ { E } + l _ { H } } _ { r, s = 1 }.$$
Then recalling the definition of Ω 1 from (75), and the functions f u l E +1 , . . . , f u l E + l H from (88)-(89), we have
$$\widehat { \Xi } _ { [ l _ { E } + 1 \colon l _ { E } + l _ { H } ], [ l _ { E } + 1 \colon l _ { E } + l _ { H } ] } ^ { 2 l _ { E } + l _ { H }, r } = \widehat { \Omega } ^ { 1, r } \implies \Xi _ { [ l _ { E } + 1 \colon l _ { E } + l _ { H } ], [ l _ { E } + 1 \colon l _ { E } + l _ { H } ] } ^ { 2 l _ { E } + l _ { H } } = \Omega ^ { 1 }.$$
Since we have shown that ( h l E +1 , . . . , h l E + l H ) = H 1 (see (86)), and Corollary A.2 states that
$$( d _ { \mathbf c } ^ { 1 }, h _ { \mathbf c } ^ { l _ { E } + 1 }, \dots, h _ { \mathbf c } ^ { l _ { E } + l _ { H } } ) \stackrel { W _ { 2 } } { \to } ( \bar { d } _ { \mathbf c } ^ { 1 }, \bar { h } ^ { l _ { E } + 1 }, \dots, \bar { h } ^ { l _ { E } + l _ { H } } ),$$
where ¯ d 1 c and ( ¯ h l E +1 , . . . , h ¯ l E + l H ) ∼ N ( 0 Ξ , l E + l H [ l E +1: l E + l H ] , l [ E +1: l E + l H ] ) are indepedent. Recalling that d 1 c = β c , by the equivalence of the covariance matrices in (97), we have
$$( H ^ { 1 } _ { \mathcal { J } _ { \epsilon } }, \beta _ { \epsilon } ) \stackrel { W _ { 2 } } { \to } ( \bar { H } ^ { 1 } _ { \epsilon }, \bar { \beta } ), \quad \text{where} \, \bar { \beta } \text{ is independent of} \, \bar { H } ^ { 1 } _ { \epsilon } \sim \mathcal { N } ( 0, \Omega ^ { 1 } ).$$
Convergence of ( E , γ 1 ) . Recall that v s c = f v s ( h , . . . , h 1 c t c , d 1 c , . . . , d L d c , c ) and ¯ v s c = f v s ( ¯ h , . . . , h 1 ¯ t , d ¯ 1 c , . . . , d ¯ L d c , c ) , for c ∈ [ C ] . Corollary A.2 implies that ( v , . . . , v 1 c 2 l E + l H c ) W 2 → (¯ v , . . . , v 1 c ¯ 2 l E + l H c ) . Moreover, we have shown in (87) that ( v l E + l H +1 , . . . , v 2 l E + l H ) = 1 √ δ ̂ H 1 = 1 √ δ ˜ ( f 1 H ,β, 1 C ) . Thus, using (98) and noting that ˜ f 1 satisfying the polynomial growth condition in (5), we have (¯ v l E + l H +1 c , . . . , v ¯ 2 l E + l H c ) = 1 √ δ ˜ ( ¯ f 1 H ,β, 1 c ¯ c ) , for c ∈ [ C ] .
Corollary states that A.2 that ( e , . . . , e 1 r 2 l E + l H r ) W 2 → (¯ e , . . . , e 1 ¯ 2 l E + l H ) ∼ N (0 , Γ 2 l E + l H ) , where
$$\Gamma ^ { 2 l _ { E } + l _ { H } } = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \widehat { \Gamma } ^ { 2 l _ { E } + l _ { H }, c }, \quad \widehat { \Gamma } ^ { 2 l _ { E } + l _ { H }, c } = \left ( \mathbb { E } \{ \bar { v } _ { c } ^ { r } \bar { v } _ { c } ^ { s } \} \right ) _ { r, s = 1 } ^ { 2 l _ { E } + l _ { H } }.$$
Then recalling the definition of Π 1 from (75), and the functions f v l E + l H +1 , . . . , f v 2 l E + l H from (91) and (93), we have
$$\widehat { \Gamma } _ { [ l _ { E } + l _ { H } + 1 ; 2 l _ { E } + l _ { H } ], [ l _ { E } + l _ { H } + 1 ; 2 l _ { E } + l _ { H } ] } ^ { 2 l _ { E } + l _ { H } } = \widehat { \Pi } ^ { 1, c } \, \Longrightarrow \, \Gamma _ { [ l _ { E } + l _ { H } + 1 ; 2 l _ { E } + l _ { H } ], [ l _ { E } + l _ { H } + 1 ; 2 l _ { E } + l _ { H } ] } ^ { 2 l _ { E } + l _ { H } } = \Pi ^ { 1 }. \quad ( 1 0 0 )$$
Since ( e l E + l H +1 , . . . , e 2 l E + l H ) = E 1 from (87), and Corollary A.2 states that
$$( c _ { r } ^ { 1 }, e _ { r } ^ { l _ { E } + l _ { H } + 1 }, \dots, e _ { r } ^ { 2 l _ { E } + l _ { H } } ) \stackrel { W _ { 2 } } { \to } ( \bar { c } _ { r } ^ { 1 }, \bar { e } ^ { l _ { E } + l _ { H } + 1 }, \dots, \bar { e } ^ { 2 l _ { E } + l _ { H } } ),$$
where ¯ c 1 r and (¯ e l E + l H +1 , . . . , e ¯ 2 l E + l H ) ∼ N ( 0 Γ , 2 l E + l H [ l E + l H +1:2 l E + l H ] , l [ E + l H +1:2 l E + l H ] ) are independent. Therefore, using c 1 r = γ r and (100), we have ( E 1 I r , : , γ r ) W 2 → ( ¯ E , γ 1 r ¯) , where ¯ E 1 r ∼ N (0 Π ) , 1 is independent of ¯ γ .
Inductive hypothesis. For k ≥ 1 , assume that we can reduce H k , ̂ R k -1 , ̂ H k , and E k to U-AMP in iterations t = l E k + l H ( k -1) + 1 , . . . , l E ( k +1) + l H k . In formulas, this means the following:
$$\text{$w=\cdot_{E}^{\prime}\nu\cdot\cdot\cdotN\vee\nolimits}^{\prime} \quad \gamma \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime}\cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \cdot_{E}^{\prime} \\ \bullet \text{For $t=l_{E}k+l_{H}(k-1)+1,\dots,h^{l_{E}k+kl_{H}$)} = H^{k}, \quad v^{l_{E}k+l_{H}(k-1)+1},\dots,v^{l_{E}k+kl_{H}=0} \\ e^{l_{E}k+l_{H}(k-1)+1},\dots,e^{l_{E}k+kl_{H}=0}, \quad ( u^{l_{E}k+l_{H}(k-1)+1},\dots, u^{l_{E}k+kl_{H}=kl_{H})} = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { k - 1 }.$$
$$\bullet & \text{ For } t = l _ { E } k + l _ { H } k + 1, \dots, l _ { E } ( k + 1 ) + l _ { H } k \colon \text{ We have} \\ & h ^ { l _ { E } k + l _ { H } k + 1 }, \dots, h ^ { l _ { E } ( k + 1 ) + l _ { H } k } = 0, \quad ( v ^ { l _ { E } k + l _ { H } k + 1 }, \dots, v ^ { l _ { E } ( k + 1 ) + l _ { H } k } ) = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { k } \\ & \quad \text{$l_{e} l_{E} k + l _ { H } k + 1 } \quad \text{$l_{E} ( k + 1 ) + l _ { H } k } \sum - \mathbb { R } ^ { k } \quad \text{$l_{E} k + l _ { H } k + 1 } \quad \text{$l_{E} ( k + 1 ) + l _ { H } k } - \mathbb { A }$$
$$h ^ { l _ { E } k + l _ { H } k + 1 }, \dots, h ^ { l _ { E } ( k + 1 ) + l _ { H } k } = 0, \quad ( v ^ { l _ { E } k + l _ { H } k + 1 }, \dots, v ^ { l _ { E } ( k + 1 ) + l _ { H } k } ) = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { k } \\ ( e ^ { l _ { E } k + l _ { H } k + 1 }, \dots, e ^ { l _ { E } ( k + 1 ) + l _ { H } k } ) = E ^ { k }, \quad u ^ { l _ { E } k + l _ { H } k + 1 }, \dots, u ^ { l _ { E } ( k + 1 ) + l _ { H } k } = 0.$$
Defining the index sets
$$\mathcal { I } _ { k } = [ l _ { E } k + l _ { H } k + 1 \, \colon l _ { E } ( k + 1 ) + l _ { H } k ], \quad \mathcal { J } _ { k } = [ l _ { E } k + l _ { H } ( k - 1 ) + 1 \, \colon l _ { E } k + l _ { H } k ],$$
we also assume that the follows convergence statements hold, for 1 ≤ s ≤ k :
$$& ( H ^ { s } _ { \mathcal { J } _ { c } ; }, \beta _ { c } ) \stackrel { W _ { 2 } } { \to } ( \bar { H } ^ { s } _ { c }, \bar { \beta } ), \quad \bar { \beta } \text{ independent of } \bar { H } ^ { s } _ { c } \sim \mathcal { N } ( 0, \Omega ^ { s } ), \\ & ( E ^ { s } _ { \mathcal { I } _ { r } ; }, \gamma _ { r } ) \stackrel { W _ { 2 } } { \to } ( \bar { E } ^ { s } _ { r }, \bar { \gamma } ), \quad \bar { \gamma } \text{ independent of } \bar { E } ^ { k } _ { r } \sim \mathcal { N } ( 0, \Pi ^ { s } ).$$
Inductive step. We need to show that we can reduce H k +1 , ̂ R k , ̂ H k +1 , and E k +1 to U-AMP in iterations t = l E ( k +1)+ l H k +1 , . . . , l E ( k +2)+ l H ( k +1) , and that the corresponding convergence statements hold. The choices for the functions ( f v t , f u t +1 ) are analogous to those in (89), (91), and the steps for reduction are very similar to the base case for t ∈ { l E + 1 , . . . , 2 L E + l H } , and are omitted for brevity. We provide the summary of the reductions below:
$$\text{omited for brevity. we province the summary of the recuions below: } \\ \bullet \text{ For } t = l _ { E } ( k + 1 ) + l _ { H } k + 1, \dots, l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) \colon \text{We have} \\ \left ( h _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } k + 1 }, \dots, h _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) } \right ) = H ^ { k + 1 }, \quad v _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } k + 1 }, \dots, v _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) } = 0 \\ \right ) _ { e ^ { l _ { E } ( k + 1 ) + l _ { H } k + 1 }, \dots, e ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) } = 0, \quad \left ( u _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } k + 1 }, \dots, u _ { E } ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) } \right ) = \frac { 1 } { \sqrt { \delta } } \widehat { R } ^ { k }. \\ \right ) _ { - } \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdclots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \cdots \quad \end{$$
$$& \bullet \text{ For } t = l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1, \dots, l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) \colon \text{ We have} \\ & \quad h ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1 }, \dots, h ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = 0, \quad \left ( v ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1 }, \dots, v ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } \right ) = \frac { 1 } { \sqrt { \delta } } \widehat { H } ^ { k + 1 } \\ & \quad \left ( e ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1 }, \dots, e ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } \right ) = E ^ { k + 1 }, \quad u ^ { l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1 }, \dots, u ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = 0.$$
$$( 1 0 5 )$$
We now show the convergence statements in Theorem A.1 for H k +1 , and E k +1 by reducing the abstract matrix-AMP SE parameters to the corresponding U-AMP SE parameters. Define the index sets
$$\mathcal { I } _ { k + 1 } = [ l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) + 1 \, \colon l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) ], \, \mathcal { J } _ { k + 1 } = [ l _ { E } ( k + 1 ) + l _ { H } k + 1 \, \colon l _ { E } ( k + 1 ) + l _ { H } ( k + 1 ) ].$$
Convergence of ( H k +1 , β ) . From (104) we have ( u l r ) l ∈J k +1 = 1 √ δ ̂ R k I r , : = 1 √ δ ˜ ( g k E k I r , : , γ r , r ) , and by the inductive hypothesis, ( E k I r , : , γ r ) W 2 → ( ¯ E ,γ k r ¯) . Since ˜ g k satisfies the polynomial growth condition in (5), Corollary A.2 implies that ( u l r ) l ∈J k +1 W 2 → (¯ ) u l r l ∈J k +1 = 1 √ δ ˜ ( ¯ g k E ,γ, k r ¯ r ) . Corollary A.2 states that ( h l c ) l ∈J k +1 W 2 → ( ¯ h l ) l ∈J k +1 ∼ N (0 Ξ , l E ( k +2)+ l H ( k +1) J k +1 , J k +1 ) , where
$$\Xi _ { \mathcal { J } _ { k + 1 }, \mathcal { J } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } \Xi _ { \mathcal { J } _ { k + 1 }, \mathcal { J } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), r }, \quad \widehat { \Xi } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), r } = \delta \left ( \mathbb { E } [ \bar { u } _ { r } \bar { u } _ { r } ^ { s } ] \right ) _ { r, s \in \mathcal { J } _ { k + 1 } }.$$
Then comparing the above with the definition of Ω k +1 in (75), we get
$$\widehat { \Xi } _ { \mathcal { J } _ { k + 1 }, \mathcal { J } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), r } = \widehat { \Omega } ^ { k + 1, r } \implies \Xi _ { \mathcal { J } _ { k + 1 }, \mathcal { J } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = \Omega ^ { k + 1 }.$$
From (104), we have ( h l ) l ∈J k +1 = H k +1 , and Corollary A.2 states that ( d , 1 c ( h l c ) l ∈J k +1 ) W 2 → ( ¯ d , 1 c ( ¯ h l ) l ∈J k +1 ) , where ¯ d 1 c and ( ¯ h l ) l ∈J k +1 are independent. Thus, by the equivalence of covariance matrices in (106) and recalling d 1 c = β c , we have ( H k +1 J c , : , β c ) W 2 → ( ¯ H k +1 c , β ¯ ) , where ¯ β is independent of ¯ H k +1 c ∼ N (0 , Ω k +1 ) .
Convergence of ( E k +1 , γ ) . From (105), we have ( v l c ) l ∈I k +1 = 1 √ δ ̂ H k +1 J c , : = 1 √ δ ˜ f k +1 ( H k +1 J c , : , β c , c ) , and we have shown that ( H k +1 J c , : , β c ) W 2 → ( ¯ H k +1 c , β ¯ ) . Since ˜ f k +1 satisfies the polynomial growth condition, Corollary A.2 implies that ( v l c ) l ∈I k +1 W 2 → (¯ ) v l c l ∈I k +1 = 1 √ δ ˜ f k +1 ( ¯ H k +1 , β, ¯ c ) .
Corollary A.2 states that ( e l r ) l ∈I k +1 W 2 → (¯ ) e l l ∈I k +1 ∼ N (0 Γ , l E ( k +2)+ l H ( k +1) I k +1 , I k +1 ) , where
$$\Gamma _ { \mathcal { I } _ { k + 1 }, \mathcal { I } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = \frac { 1 } { \mathbb { C } } \sum _ { c = 1 } ^ { C } \widehat { \Gamma } _ { \mathcal { I } _ { k + 1 }, \mathcal { I } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), \epsilon }, \quad \widehat { \Gamma } _ { \mathcal { I } _ { k + 1 }, \mathcal { I } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), \epsilon } = \left ( \mathbb { E } [ \bar { v } _ { c } ^ { r } \bar { v } _ { c } ] \right ) _ { r, s \in \mathcal { I } _ { k + 1 } }.$$
Then comparing the above with the definition of Π k +1 in (75), we get
$$\widehat { \Gamma } _ { \mathcal { I } _ { k + 1 }, \mathcal { I } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ), \epsilon } = \widehat { \Pi } ^ { k + 1, \epsilon } \implies \Gamma _ { \mathcal { I } _ { k + 1 }, \mathcal { I } _ { k + 1 } } ^ { l _ { E } ( k + 2 ) + l _ { H } ( k + 1 ) } = \Pi ^ { k + 1 }$$
From (105), we have ( e l ) l ∈I k +1 = E k +1 and Corollary A.2 states that ( c 1 , ( e l r ) l ∈I k +1 ) W 2 → ( ¯ c 1 , (¯ ) e l l ∈I k +1 ) , where ¯ c 1 and (¯ ) e l l ∈I k +1 are independent. Thus, by the equivalence of covariance matrices in (107) and recalling c 1 r = γ r , we have ( E k +1 I r , : , γ r ) W 2 → ( ¯ E k +1 r , γ ¯) , where ¯ γ is independent of ¯ E k +1 r ∼ N (0 Π , k +1 ) . This completes the proof of the inductive step, and hence, of Theorem A.1.
## A.3 Proof of Corollary A.2
The abstract AMP recursion in (79), but without the block-wise dependence of the functions f v t and f u t +1 , was analyzed in [36]. We show how the block-wise dependence can be included without loss of generality, and thereby prove Corollary A.2 by appealing to the state evolution result of [36].
The abstract AMP iteration for a generalized white noise matrix ˜ X ∈ R n × p analyzed in [36] is as follows. Given an initializer u 1 ∈ R n , side information c 1 , . . . , c L ∗ c ∈ R n and d , . . . , d 1 L ∗ d ∈ R p , all independent of ˜ X , the iterates of the abstract AMP recursion are computed as:
$$\text{prename or } a, \, \text{and are always or in a conversion } a \, \text{compute as}. \\ h ^ { t } & = \sqrt { \delta } \widetilde { X } ^ { \top } u ^ { t } - \sum _ { s = 1 } ^ { t - 1 } b _ { s } ^ { t } v ^ { s }, \quad v ^ { t } = f _ { t } ^ { v } ( h ^ { 1 }, \dots, h ^ { t }, d ^ { 1 }, \dots, d ^ { L _ { d } ^ { d } } ), \\ e ^ { t } & = \sqrt { \delta } \widetilde { X } v ^ { t } - \sum _ { s = 1 } ^ { t } a _ { s } ^ { t } u ^ { s }, \quad u ^ { t + 1 } = f _ { t + 1 } ^ { u } ( e ^ { 1 }, \dots, e ^ { t }, c ^ { 1 }, \dots, c ^ { L _ { c } ^ { d } } ), \\ \text{and are } a, \, \text{and are } v \,. \, \text{$\mathfrak{d}$} + L _ { d } ^ { \prime } \cdots \, \mathfrak{D} \, \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} + \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \text{$\mathfrak{d}$} \,. \, \end{$$
where the functions f v t : R t + L ∗ d → R , f u t +1 : R t + L ∗ c → R act row-wise. The memory coefficients { b t s } s<t and { a t s } s ≤ t are defined below in (109). We have the following assumptions:
(E1) When n, p →∞ , we have n/p = δ > 0 , for fixed L ∗ c and L ∗ d . Furthermore, we have
$$( u ^ { 1 }, c ^ { 1 }, \dots, c ^ { L _ { c } ^ { * } } ) \stackrel { W } { \to } ( \bar { u } ^ { 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } } ) \text{ and } ( d ^ { 1 }, \dots, d ^ { L _ { d } ^ { * } } ) \stackrel { W } { \to } ( \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } } ),$$
for joint limit laws (¯ u , c 1 ¯ 1 , . . . , c ¯ L ∗ c ) and ( ¯ d , . . . , d 1 ¯ L ∗ d ) having finite moments of all orders, where E [(¯ u 1 ) 2 ] ≥ 0 . Multivariate polynomials are dense in the real L 2 -spaces of functions f : R L ∗ c +1 → R and g : R L ∗ d → R with the inner products
$$\langle f, \tilde { f } \rangle \coloneqq \mathbb { E } [ f ( \bar { u } ^ { 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } } ) \tilde { f } ( \bar { u } ^ { 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } } ) ] \text{ and } \langle g, \tilde { g } \rangle \coloneqq \mathbb { E } [ g ( \bar { u } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } } ) \tilde { g } ( \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } } ) ].$$
(E2) , (E3) , (E4) These are identical to (D2) , (D3) , (D4) , with L ∗ d replacing L d .
The state evolution covariance matrices Ξ t , Γ t ∈ R t × t are iteratively defined as follows, starting from Ξ 1 = δ E [(¯ u 1 ) 2 ] ∈ R 1 × 1 . Given Ξ t , for t ≥ 1 , let ( ¯ h , . . . , h 1 ¯ ) t ∼ N (0 Ξ ) , t independent of ( ¯ d , . . . , d 1 ¯ L ∗ d ) and define
$$\bar { v } ^ { s } = f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { s }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } } ), \ \ s \in [ t ].$$
Then, Γ t = ( E [¯ v v r ¯ ]) s t r,s =1 ∈ R t × t . Next, let (¯ e , . . . , e 1 ¯ ) t ∼ N (0 Γ ) , t independent of (¯ u , c 1 ¯ 1 , . . . , ¯ c L ∗ c ) and define
$$\bar { u } ^ { s + 1 } = f _ { s + 1 } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { s }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } } ), \ \ s \in [ t ].$$
Then, Ξ t +1 = ( δ · E [¯ u u r ¯ ]) s t +1 r,s =1 ∈ R ( +1) t × ( +1) t . The memory coefficients in (108) are then defined as
$$a _ { s } ^ { t } = \mathbb { E } [ \partial _ { s } f _ { t } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } } ) ] \text{ and } b _ { s } ^ { t } = \delta \cdot \mathbb { E } [ \partial _ { s } f _ { t } ^ { u } ( \bar { e } ^ { 1 } \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } } ) ],$$
where ∂ s denotes partial derivative in the s th argument. The following theorem gives the state evolution result for the abstract AMP recursion.
Theorem A.3. [36, Theorem 2.17] Let ˜ X ∈ R n × p be a generalized white noise matrix (as defined in Definition 5.1) with variance profile S ∈ R n × p , and let u , c 1 1 , . . . , c L ∗ c , d 1 , . . . , d L ∗ d be independent of ˜ X and satisfy Assumptions (E1)-(E4). Further assume that each matrix Ξ t and Γ t is non-singular. Then for any fixed t ≥ 1 , almost surely as n, p → ∞ with n/p = δ ∈ (0 , ∞ ) , the iterates of the abstract AMP in (108) satisfy
$$& ( u ^ { 1 }, c ^ { 1 }, \dots, c ^ { L _ { c } ^ { * } }, e ^ { 1 }, \dots, e ^ { t } ) \stackrel { W _ { 2 } } { \to } ( \bar { u } ^ { 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } ^ { * } }, \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t } ) \\ & ( d ^ { 1 }, \dots, d ^ { L _ { d } ^ { * } }, h ^ { 1 }, \dots, h ^ { t } ) \stackrel { W _ { 2 } } { \to } ( \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } ^ { * } }, \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t } ),$$
where ( ¯ h , . . . , h 1 ¯ ) t ∼ N (0 , Ξ ) t and (¯ e , . . . , e 1 ¯ ) t ∼ N (0 Γ ) , t are independent of (¯ u , c 1 ¯ 1 . . . . , c ¯ L ∗ c ) and ( ¯ d , . . . , d 1 ¯ L ∗ d ) .
To obtain the U-AMP recursion (79) from the abstract AMP recursion in (108), we choose L ∗ c = L c +1 and L ∗ d = L d +1 , and the side information vectors d L d +1 ∈ R p and c L c +1 ∈ R n are set as
$$d ^ { L _ { d } + 1 } = \mathcal { C }, \quad c ^ { L _ { c } + 1 } = \mathcal { R }.$$
The functions f v t and f u t +1 , as well as the initializer u 1 are the same as those in (108).
The memory coefficients in (109) can then be expressed as:
With this choice, the empirical distribution of d L d +1 converges to ¯ d L d +1 ∼ Uniform ([ C ]) , and the empirical distribution of c L c +1 converges to ¯ c L c +1 ∼ Uniform ([ R ]) . Moreover, Assumption (E1) is equivalent to Assumption (D1) of Corollary A.2. To see this, for r ∈ [ R ] , let (¯ u , c 1 r ¯ 1 r , . . . , c ¯ L c r ) be random variables whose joint law equals the conditional law of (¯ u , c 1 ¯ 1 , . . . , c ¯ L c ) given ¯ c L c +1 = r . Similarly, for c ∈ [ C ] , let ( ¯ d , . . . , d 1 c ¯ L d c ) be jointly distributed according to the conditional law of ( ¯ d , . . . , d 1 ¯ L d ) given ¯ d L d +1 = c .
$$\text{$1$} \text{ memory connections in (10)$} \text{ can even we expressed as} \colon \\ a ^ { t } _ { s } = \mathbb { E } [ \partial _ { s } f ^ { v } _ { t } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { h } ^ { 1 }, \dots, \bar { d } ^ { L d + 1 } ) ] & = \mathbb { E } \left [ \mathbb { E } [ \partial _ { s } f ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L d + 1 } ) \, | \, \bar { d } ^ { L d + 1 } ] \right ] \\ & = \frac { 1 } { \mathbb { C } } \sum _ { c = 1 } ^ { c } \mathbb { E } [ \partial _ { s } f ^ { v } _ { t } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L d }, c ) \, | \, \bar { d } ^ { L d + 1 } = c ], \\ & = \frac { 1 } { \mathbb { C } } \sum _ { c = 1 } ^ { c } \mathbb { E } [ \partial _ { s } f ^ { v } _ { t } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 } _ { c }, \dots, \bar { d } ^ { L d } _ { c }, c ) ]. \\ \text{throw for that last enough if you need to set that } \bar { h } ^ { 1 } _ { t } \, \text{$\bar{ }h } ^ { 1 } \, \text{$\bar{ }i$} \, \text{information} \, of $f(\bar{ }l_{1}$} } \quad \bar { h } ^ { L d + 1 } \, \text{$\bar{ }l_{1}$} \, \text{$Similarity}$$
where for the last equality we used fact that ( ¯ h , . . . , h 1 ¯ ) t is independent of ( ¯ d , . . . , d 1 ¯ L d +1 ) . Similarly, we have
$$b _ { s } ^ { t } = \delta \mathbb { E } [ \partial _ { s } f _ { t } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } + 1 } ) ] & = \delta \mathbb { E } \left [ \mathbb { E } [ \partial _ { s } f _ { t } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L _ { c } + 1 } ) \, | \, \bar { c } ^ { L _ { c } + 1 } ] \right ] \\ & = \frac { \delta } { R } \sum _ { r = 1 } ^ { R } \mathbb { E } [ \partial _ { s } f _ { t } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 } _ { r }, \dots, \bar { c } ^ { L _ { c } + 1 } _ { r } ) ].$$
$$\begin{smallmatrix} 1 \end{smallmatrix}$$
Next, for r, s ∈ [ t ] , the ( r, s ) th element of the state evolution matrix Γ t ∈ R t × t is
$$\left ( \Gamma ^ { t } \right ) _ { r, s } & = \mathbb { E } \left [ f _ { r } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } + 1 } ) f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } + 1 } ) \right ] \\ & = \mathbb { E } \left [ \mathbb { E } \left [ f _ { r } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } + 1 } ) f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 }, \dots, \bar { d } ^ { L _ { d } + 1 } ) \left | \bar { d } ^ { L _ { d } + 1 } \right ] \right ] \\ & = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \mathbb { E } \left [ f _ { r } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 } _ { c }, \dots, \bar { d } ^ { L _ { d } } _ { c }, c ) f _ { s } ^ { v } ( \bar { h } ^ { 1 }, \dots, \bar { h } ^ { t }, \bar { d } ^ { 1 } _ { c }, \dots, \bar { d } ^ { L _ { d } } _ { c }, c ) \right ] = \frac { 1 } { C } \sum _ { c = 1 } ^ { C } \mathbb { E } \left [ v _ { c } ^ { v } v _ { c } ^ { c } \right ], \\ \cdot \quad \ \ - \quad \ \frac { \sim } { c } \sum _ { c = 1 } ^ { C } \frac { \sim } { c } \sum _ { c = 1 } ^ { C } \frac { \sim } { c } \sum _ { c = 1 } ^ { C } \frac { \sim } { c } \sum _ { c = 1 } ^ { C } \frac { \sim } { c } \cdot \quad \ \.$$
where ¯ v s c = f v s ( ¯ h , . . . , h 1 ¯ t , d ¯ 1 c , . . . , d ¯ L d c , c ) . Similarly, we have
$$( \Xi ^ { t } ) _ { r, s } = \frac { \delta } { R } \sum _ { r = 1 } ^ { R } \mathbb { E } [ f _ { r } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L c + 1 } = r ) f _ { s } ^ { u } ( \bar { e } ^ { 1 }, \dots, \bar { e } ^ { t - 1 }, \bar { c } ^ { 1 }, \dots, \bar { c } ^ { L c + 1 } = r ) ] = \frac { \delta } { R } \sum _ { r = 1 } ^ { R } \mathbb { E } [ \bar { u } ^ { r } _ { r } \bar { u } ^ { s } _ { r } ],$$
where for s ≥ 1 , we have ¯ u s +1 r = f u s +1 (¯ e , . . . , e 1 ¯ s , c ¯ 1 r , . . . , c ¯ L c r , r ) . We have shown that with the choice of side information in (110), the AMP recursion (108) matches with that in (79), and the corresponding state evolution recursions also match. Applying Theorem A.3 and recalling the definitions of R C , from (51) gives us Corollary A.2.
## B Proof of Lemma 3.6
For δ > 0 , evaluating U b ( 0 ; δ ) in (31) with b 0 := δσ 2 gives
$$U ( b _ { 0 } ; \delta ) = - \frac { \delta } { 2 } + \delta \log 2 + 2 I \left ( \bar { \beta }, \sqrt { \frac { 1 } { 2 \sigma ^ { 2 } } } \bar { \beta } + G \right ).$$
We pause to state the following auxiliary result.
Lemma B.1. [34, Proposition 7.15] For a discrete distribution P ¯ β with finite alphabet, we have
$$\lim _ { s \to \infty } \frac { I ( \bar { \beta } ; \sqrt { s } \bar { \beta } + G ) } { \frac { 1 } { 2 } \log s } = 0.$$
Lemma B.1 implies that for any ∆ > 0 , we have I ( ¯ ; β √ sβ ¯ + G ) ≤ ∆ 2 log s for all sufficiently large s . Taking s = 1 2 σ 2 further implies that for sufficiently small σ , we have
$$I \left ( \bar { \beta } ; \sqrt { \frac { 1 } { 2 \sigma ^ { 2 } } } \bar { \beta } + G \right ) \leq \frac { \Delta } { 2 } \log \left ( \frac { 1 } { 2 \sigma ^ { 2 } } \right ) \\ \Longrightarrow & U ( b _ { 0 } ; \delta ) \leq - \frac { \delta } { 2 } + \delta \log 2 + \Delta \log \left ( \frac { 1 } { 2 \sigma ^ { 2 } } \right ).$$
Hence, for any ∆ > 0 , there exists σ 0 (∆) > 0 such that for all σ < σ 0 (∆) we have the following for all b ∈ (0 , Var( ¯ )] β and δ > 0 :
$$\dots \nu \subset \{ \nu, \dots \nu \nmid \nu \nmid \nu \times \nu. \\ U ( b ; \delta ) - U ( b _ { 0 } ; \delta ) & \geq - \delta \left ( 1 - \frac { \delta \sigma ^ { 2 } } { b + \delta \sigma ^ { 2 } } \right ) + \delta \log \left ( 1 + \frac { b } { \delta \sigma ^ { 2 } } \right ) + 2 I \left ( \tilde { \beta } ; \sqrt { \frac { \delta } { b + \delta \sigma ^ { 2 } } } \tilde { \beta } + G \right ) \\ & + \frac { \delta } { 2 } - \delta \log 2 - \Delta \log \left ( \frac { 1 } { 2 \sigma ^ { 2 } } \right ) \\ & \geq \delta \log \left ( \frac { b } { \delta \sigma ^ { 2 } } \right ) - \Delta \log \left ( \frac { 1 } { 2 \sigma ^ { 2 } } \right ) - \frac { \delta } { 2 } - \delta \log 2, \\ \dots \dots \nmid \nu \dots \nmid \nu \nmid \dots \nmid \nu \nmid \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \dots \nmid \nu \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \ delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \dta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \Delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta \delta$$
where (a) uses (31) and (113), and (b) uses the non-negativity of mutual information. Now, for σ < σ 0 (∆) and
$$b > 2 \delta e ^ { \frac { 1 } { 2 } } \left ( \frac { 1 } { 2 } \right ) ^ { \frac { \Delta } { \delta } } \sigma ^ { 2 - \frac { 2 \Delta } { \delta } },$$
the lower bound in (114) is strictly positive. Therefore, U b δ ( ; ) -U b ( 0 ; δ ) > 0 for b satisfying (115), implying that these values of b cannot be minimizers of U b δ ( ; ) . Therefore, for any δ, ∆ > 0 and σ < σ 0 (∆) , we have:
$$\max \left \{ \underset { b \in ( 0, \text{Var} ( \bar { \beta } ) ] } { \arg \min } U ( b ; \delta ) \right \} \leq 2 \delta e ^ { \frac { 1 } { 2 } } \left ( \frac { 1 } { 2 } \right ) ^ { \frac { \Delta } { \delta } } \sigma ^ { 2 - \frac { 2 \Delta } { \delta } } < \frac { 7 } { 2 } \delta ( \sigma ^ { 2 - \frac { 2 \Delta } { \delta } } ).$$
## C Implementation Details
SC-AMP denoiser and state evolution parameters for QGT. The Bayes-optimal denoiser f k +1 in (19) can be computed using the prior ¯ β ∼ Bernoulli ( π ) . For j ∈ J c and k ≥ 1 , we have
$$f _ { k } ( s, c ) & = \mathbb { E } [ \bar { \beta } \, | \, ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G = s ] = \frac { \mathbb { P } \left [ \bar { \beta } = 1 \right ] \cdot \mathbb { P } \left [ ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G = s | \bar { \beta } = 1 \right ] } { \sum _ { \bar { \beta } \in \{ 0, 1 \} } \mathbb { P } [ \bar { \beta } ] \cdot \mathbb { P } [ ( \chi _ { c } ^ { k } ) ^ { 2 } \bar { \beta } + \chi _ { c } ^ { k } G = s | \bar { \beta } ] } \\ & = \frac { \pi \phi ( ( s - ( \chi _ { c } ^ { k } ) ^ { 2 } ) / \chi _ { c } ^ { k } ) } { \pi \phi ( ( s - ( \chi _ { c } ^ { k } ) ^ { 2 } ) / \chi _ { c } ^ { k } ) },$$
where ϕ x ( ) is the standard normal density. Instead of precomputing the state evolution parameters ( χ k c ) , they can be estimated from the SC-AMP iterates as:
$$\left ( \widehat { \chi } _ { c } ^ { k } \right ) ^ { 2 } = \sum _ { r = 1 } ^ { R } \widetilde { W } _ { r c } \left ( \left \| \widetilde { \Theta } _ { r } ^ { k } \right \| _ { 2 } ^ { 2 } \right ) ^ { - 1 } \quad \text{ for } c \in [ C ],$$
where ˜ Θ k r = (Θ ) ˜ k i i ∈I r is the restriction of ˜ Θ k to indices I r . The derivative ∂ f 1 k ( β , k j c ) , required for b k in (15), can be obtained by applying the Quotient rule to the last expression in (116).
SC-AMP denoiser and state evolution parameters for pooled data. The Bayes-optimal denoiser f k : R L × R → R L in the SC-AMP algorithm in (40) is computed as:
$$f _ { k } ( s, c ) & = \mathbb { E } [ \bar { B } \, | \, \bar { B } + G _ { \epsilon } ^ { k } = s ] = \sum _ { l = 1 } ^ { L } e _ { l } \frac { \mathbb { P } [ \bar { B } = e _ { l } ] \mathbb { P } [ \bar { B } + G _ { \epsilon } ^ { k } = s | \bar { B } = e _ { l } ] } { \mathbb { P } [ \bar { B } + G _ { \epsilon } ^ { k } = s ] } \\ & = \frac { ( a ) } { \sum _ { l = 1 } ^ { L } \pi _ { l } e _ { l } \exp \left ( - \frac { 1 } { 2 } ( e _ { l } - s ) ^ { \intercal } \left ( \mathbb { T } _ { \epsilon } ^ { k } \right ) ^ { - 1 } ( e _ { l } - s ) \right ) } } { \sum _ { l = 1 } ^ { L } \pi _ { l } \exp \left ( - \frac { 1 } { 2 } ( e _ { l } - s ) ^ { \intercal } \left ( \mathbb { T } _ { \epsilon } ^ { k } \right ) ^ { - 1 } ( e _ { l } - s ) \right ) }, \\ & \quad \quad \quad$$
where (a) follows since G k c ∼ N ( 0 T , k c ) . The state evolution parameters ϕ k r for r ∈ [ R ] are estimated from matrix SC-AMP iterates as
$$\hat { \phi } _ { r } ^ { k } = \frac { 1 } { n / R } \sum _ { i \in \mathcal { I } _ { r } } ( \widetilde { \Theta } _ { i } ^ { k } ) ^ { \top } \widetilde { \Theta } _ { i } ^ { k }, \quad \text{ for } r \in [ R ].$$
The Jacobian f ′ k ( B k j, : , c ) in (42) can be computed for all j ∈ [ p ] can be computed by applying the Quotient rule to (118), following the method in [14, App. D.1].
Potential function. To generate the curves in Figure 2, for each δ , the potential function U b δ ( ; ) in in (31) is evaluated at 500 data points between 0 and Var ( ¯ ) β (for π = 0 1 . , Var ( ¯ ) = β π -π 2 = 0 09 . ). To analyze the noiseless QGT model, we set σ = 1 × 10 -30 to avoid computational instability. The mutual information term in (31) is computed via numerical integration (instead of Monte Carlo methods) to ensure that the curves are smooth.
## References
| [1] | M. Aldridge, O. Johnson, and J. Scarlett, 'Group testing: an information theory perspective,' Founda- tions and Trends® in Communications and Information Theory , vol. 15, no. 3-4, pp. 196-392, 2019. |
|-------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [2] | C. L. Chan, S. Jaggi, V. Saligrama, and S. Agnihotri, 'Non-adaptive group testing: Explicit bounds and novel algorithms,' IEEE Transactions on Information Theory , vol. 60, no. 5, pp. 3019-3035, 2014. |
| [3] | O. Gebhard, O. Johnson, P. Loick, and M. Rolvien, 'Improved bounds for noisy group testing with constant tests per item,' IEEE Transactions on Information Theory , vol. 68, no. 4, pp. 2604-2621, 2021. |
| [4] | O. Gebhard, M. Hahn-Klimroth, O. Parczyk, M. Penschuck, M. Rolvien, J. Scarlett, and N. Tan, 'Near- optimal sparsity-constrained group testing: Improved bounds and algorithms,' IEEE Transactions on Information Theory , vol. 68, no. 5, pp. 3253-3280, 2022. |
| [5] | N. Tan, W. Tan, and J. Scarlett, 'Performance bounds for group testing with doubly-regular designs,' IEEE Transactions on Information Theory , vol. 69, no. 2, pp. 1224-1243, 2023. |
| [6] | N. H. Bshouty, 'Optimal algorithms for the coin weighing problem with a spring scale,' Conference on Learning Theory , 2009. |
| [7] | C.-C. Cao, C. Li, and X. Sun, 'Quantitative group testing-based overlapping pool sequencing to identify rare variant carriers,' BMC Bioinformatics , vol. 15, no. 1, pp. 1-14, 2014. |
| [8] | G. De Marco, T. Jurdziński, and D. R. Kowalski, 'Optimal channel utilization with limited feedback,' Journal of Computer and System Sciences , vol. 119, pp. 21-33, 2021. |
| [9] | C. Wang, Q. Zhao, and C.-N. Chuah, 'Group testing under sum observations for heavy hitter detection,' Information Theory and Applications Workshop , pp. 149-153, 2015. |
| [10] | I.-H. Wang, S.-L. Huang, K.-Y. Lee, and K.-C. Chen, 'Data extraction via histogram and arithmetic mean queries: Fundamental limits and algorithms,' IEEE International Symposium on Information Theory , pp. 1386-1390, 2016. |
| [11] | J. Scarlett and V. Cevher, 'Phase transitions in the pooled data problem,' Advances in Neural Infor- mation Processing Systems , vol. 30, 2017. |
| [12] | V. Grebinski and G. Kucherov, 'Optimal reconstruction of graphs under the additive model,' Algorith- mica , vol. 28, pp. 104-124, 2000. |
| [13] | A. El Alaoui, A. Ramdas, F. Krzakala, L. Zdeborová, and M. I. Jordan, 'Decoding from pooled data: Sharp information-theoretic bounds,' SIAM Journal on Mathematics of Data Science , vol. 1, no. 1, pp. 161-188, 2019. |
| [14] | N. Tan, P. Pascual Cobo, J. Scarlett, and R. Venkataramanan, 'Approximate message passing with rigorous guarantees for pooled data and quantitative group testing,' to appear in SIAM Journal on Mathematics of Data Science , 2024. arXiv:2309.15507. |
| [15] | Y. Kabashima, 'A CDMA multiuser detection algorithm on the basis of belief propagation,' Journal of Physics A: Mathematical and General , vol. 36, no. 43, pp. 11 111-11 121, Oct 2003. |
| [16] | M. Bayati and A. Montanari, 'The dynamics of message passing on dense graphs, with applications to compressed sensing,' IEEE Transactions on Information Theory , vol. 57, pp. 764-785, 2011. |
| [17] | D. L. Donoho, A. Maleki, and A. Montanari, 'Message passing algorithms for compressed sensing,' Proceedings of the National Academy of Sciences , vol. 106, pp. 18 914-18 919, 2009. |
| [18] | F. Krzakala, M. Mézard, F. Sausset, Y. Sun, and L. Zdeborová, 'Probabilistic reconstruction in com- pressed sensing: algorithms, phase diagrams, and threshold achieving matrices,' Journal of Statistical Mechanics: Theory and Experiment , vol. 2012, no. 8, 2012. |
| [19] | S. Rangan, 'Generalized approximate message passing for estimation with random linear mixing,' IEEE International Symposium on Information Theory , 2011. |
|--------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [20] | J. Ma, J. Xu, and A. Maleki, 'Optimization-based AMP for phase retrieval: The impact of initialization and ℓ 2 regularization,' IEEE Transactions on Information Theory , vol. 65, no. 6, pp. 3600-3629, 2019. |
| [21] | A. Maillard, B. Loureiro, F. Krzakala, and L. Zdeborová, 'Phase retrieval in high dimensions: Statistical and computational phase transitions,' in Neural Information Processing Systems , 2020. |
| [22] | M. Mondelli and R. Venkataramanan, 'Approximate message passing with spectral initialization for generalized linear models,' International Conference on Artificial Intelligence and Statistics , pp. 397- 405, 2021. |
| [23] | N. Tan and R. Venkataramanan, 'Mixed regression via approximate message passing,' Journal of Ma- chine Learning Research , vol. 24, pp. 1-44, 2023. |
| [24] | Y. Deshpande and A. Montanari, 'Information-theoretically optimal sparse PCA,' in IEEE International Symposium on Information Theory , 2014, pp. 2197-2201. |
| [25] | A. K. Fletcher and S. Rangan, 'Iterative reconstruction of rank-one matrices in noise,' Information and Inference: A Journal of the IMA , vol. 7, no. 3, pp. 531-562, 2018. |
| [26] | Y. Kabashima, F. Krzakala, M. Mézard, A. Sakata, and L. Zdeborová, 'Phase transitions and sample complexity in Bayes-optimal matrix factorization,' IEEE Transactions on Information Theory , vol. 62, no. 7, pp. 4228-4265, 2016. |
| [27] | T. Lesieur, F. Krzakala, and L. Zdeborová, 'Constrained low-rank matrix estimation: Phase transi- tions, approximate message passing and applications,' Journal of Statistical Mechanics: Theory and Experiment , vol. 2017, no. 7, p. 073403, 2017. |
| [28] | A. Montanari and R. Venkataramanan, 'Estimation of low-rank matrices via approximate message passing,' Annals of Statistics , vol. 45, no. 1, pp. 321-345, 2021. |
| [29] | G. Li, W. Fan, and Y. Wei, 'Approximate message passing from random initialization with applica- tions to Z 2 synchronization,' Proceedings of the National Academy of Sciences , vol. 120, no. 31, p. e2302930120, 2023. |
| [30] | R. Rossetti and G. Reeves, 'Approximate message passing for the matrix tensor product model,' 2023, arXiv: 2306.15580. |
| [31] | O. Y. Feng, R. Venkataramanan, C. Rush, and R. J. Samworth, 'A unifying tutorial on approximate message passing,' Foundations and Trends in Machine Learning , 2022. |
| [32] | S. Kudekar and H. D. Pfister, 'The effect of spatial coupling on compressive sensing,' in 48th Annual Allerton Conf. Commun., Control, and Computing , 2010. |
| [33] | K. Takeuchi, T. Tanaka, and T. Kawabata, 'Performance improvement of iterative multiuser detec- tion for large sparsely spread CDMA systems by spatial coupling,' IEEE Transactions on Information Theory , vol. 61, no. 4, pp. 1768-1794, 2015. |
| [34] | D. L. Donoho, A. Javanmard, and A. Montanari, 'Information-theoretically optimal compressed sensing via spatial coupling and approximate message passing,' IEEE Transactions on Information Theory vol. 59, no. 11, pp. 7434-7464, Nov. 2013. |
| [35] | P. Pascual Cobo, K. Hsieh, and R. Venkataramanan, 'Bayes-optimal estimation in generalized linear models via spatial coupling,' in IEEE International Symposium on Information Theory (ISIT) , 2023, arXiv:2309.08404. |
| [36] | T. Wang, X. Zhong, and Z. Fan, 'Universality of approximate message passing algorithms and tensor networks,' 2022, to appear in Annals of Applied Probability . arXiv:2206.13037. |
| [37] | A. Yedla, Y.-Y. Jian, P. S. Nguyen, and H. D. Pfister, 'A simple proof of Maxwell saturation for coupled scalar recursions,' IEEE Transactions on Information Theory , vol. 60, no. 11, pp. 6943-6965, 2014. |
|--------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| [38] | A. J. Felstrom and K. S. Zigangirov, 'Time-varying periodic convolutional codes with low-density parity- check matrix,' IEEE Transactions on Information Theory , vol. 45, no. 6, pp. 2181-2191, Sept. 1999. |
| [39] | S. Kudekar, T. Richardson, and R. L. Urbanke, 'Spatially coupled ensembles universally achieve capacity under belief propagation,' IEEE Transactions on Information Theory , vol. 59, no. 12, pp. 7761-7813, Dec 2013. |
| [40] | A. Coja-Oghlan, O. Gebhard, M. Hahn-Klimroth, and P. Loick, 'Optimal group testing,' Conference on Learning Theory , pp. 1374-1388, 2020. |
| [41] | A. Coja-Oghlan, M. Hahn-Klimroth, L. Hintze, D. Kaaser, L. Krieg, M. Rolvien, and O. Scheftelowitsch, 'Noisy group testing via spatial coupling,' 2024, arXiv:2402.02895. |
| [42] | M. M. Mashauri, A. G. i Amat, and M. Lentmaier, 'Low-density parity-check codes and spatial coupling for quantitative group testing,' in IEEE International Symposium on Information Theory (ISIT) , 2023. |
| [43] | --, 'Threshold saturation for quantitative group testing with low-density parity-check codes,' in IEEE International Symposium on Information Theory (ISIT) , 2024. |
| [44] | E. Karimi, F. Kazemi, A. Heidarzadeh, K. R. Narayanan, and A. Sprintson, 'Sparse graph codes for non-adaptive quantitative group testing,' in IEEE Information Theory Workshop , 2019. |
| [45] | --, 'Non-adaptive quantitative group testing using irregular sparse graph codes,' in Annual Allerton Conference on Communication, Control, and Computing , 2019. |
| [46] | M. Hahn-Klimroth and N. Müller, 'Near optimal efficient decoding from pooled data,' Conference on Learning Theory , vol. 178, pp. 3395-3409, 2022. |
| [47] | O. Gebhard, M. Hahn-Klimroth, D. Kaaser, and P. Loick, 'Information-theoretic and algorithmic as- pects of parallel and distributed reconstruction from pooled data,' Journal of Parallel and Distributed Computing , vol. 180, p. 104718, 2023. |
| [48] | M. Soleymani and T. Javidi, 'A non-adaptive algorithm for the quantitative group testing problem,' in Proceedings of Thirty Seventh Conference on Learning Theory , 2024. |
| [49] | M. Hahn-Klimroth and D. Kaaser, 'Distributed reconstruction of noisy pooled data,' IEEE 42nd In- ternational Conference on Distributed Computing Systems , pp. 89-99, 2022. |
| [50] | Y.-H. Li and I.-H. Wang, 'Combinatorial quantitative group testing with adversarially perturbed mea- surements,' IEEE Information Theory Workshop , pp. 1-5, 2021. |
| [51] | M. Hahn-Klimroth, D. Kaaser, and M. Rau, 'Efficient approximate recovery from pooled data using doubly regular pooling scheme,' 2023, arXiv:2303.00043. |
| [52] | M. Soleymani and T. Javidi, 'Quantitative group testing with tunable adaptation,' in IEEE Interna- tional Symposium on Information Theory , 2024. |
| [53] | C. Rush, K. Hsieh, and R. Venkataramanan, 'Capacity-achieving spatially coupled sparse superposition codes with AMP decoding,' IEEE Transactions on Information Theory , vol. 67, no. 7, pp. 4446-4484, 2021. |
| [54] | G. Reeves and H. D. Pfister, 'The replica-symmetric prediction for random linear estimation with Gaussian matrices is exact,' IEEE Transactions on Information Theory , vol. 65, no. 4, pp. 2252-2283, 2019. |
| [55] | J. Barbier, N. Macris, M. Dia, and F. Krzakala, 'Mutual information and optimality of approximate message-passing in random linear estimation,' IEEE Transactions on Information Theory , vol. 66, no. 7, |
- [56] P. Pascual Cobo, 'Code for spatially coupled AMP for pooled data and quantitative group testing,' https://github.com/PabloPasc/SC\_AMP\_QGT, 2024.
- [57] S. Foucart and H. Rauhut, An invitation to compressive sensing . Springer, 2013.
- [58] X. Liu, P. Pascual Cobo, and R. Venkataramanan, 'Many-user multiple access with random user activity,' in IEEE International Symposium on Information Theory (ISIT) , 2024.
- [59] R. Vershynin, High-dimensional probability: An introduction with applications in data science . Cambridge University Press, 2018, vol. 47.
- [60] Z. Bao, Q. Han, and X. Xu, 'A leave-one-out approach to approximate message passing,' 2023, arXiv:2312.05911.
- [61] R. Berthier, A. Montanari, and P.-M. Nguyen, 'State evolution for approximate message passing with non-separable functions,' Information and Inference: A Journal of the IMA , vol. 9, no. 1, pp. 33-79, 2020.
- [62] P. Nikolopoulos, S. R. Srinivasavaradhan, T. Guo, C. Fragouli, and S. N. Diggavi, 'Community-aware group testing,' IEEE Transactions on Information Theory , vol. 69, no. 7, pp. 4361-4383, 2023.
- [63] S. Ahn, W.-N. Chen, and A. Özgür, 'Adaptive group testing on networks with community structure: The stochastic block model,' IEEE Transactions on Information Theory , vol. 69, no. 7, pp. 4758-4776, 2023. | null | [
"Nelvin Tan",
"Pablo Pascual Cobo",
"Ramji Venkataramanan"
] | 2024-08-01T08:48:26+00:00 | 2024-08-01T08:48:26+00:00 | [
"cs.IT",
"eess.SP",
"math.IT"
] | Quantitative Group Testing and Pooled Data in the Linear Regime with Sublinear Tests | In the pooled data problem, the goal is to identify the categories associated
with a large collection of items via a sequence of pooled tests. Each pooled
test reveals the number of items in the pool belonging to each category. A
prominent special case is quantitative group testing (QGT), which is the case
of pooled data with two categories. We consider these problems in the
non-adaptive and linear regime, where the fraction of items in each category is
of constant order. We propose a scheme with a spatially coupled Bernoulli test
matrix and an efficient approximate message passing (AMP) algorithm for
recovery. We rigorously characterize its asymptotic performance in both the
noiseless and noisy settings, and prove that in the noiseless case, the AMP
algorithm achieves almost-exact recovery with a number of tests sublinear in
the number of items. For both QGT and pooled data, this is the first efficient
scheme that provably achieves recovery in the linear regime with a sublinear
number of tests, with performance degrading gracefully in the presence of
noise. Numerical simulations illustrate the benefits of the spatially coupled
scheme at finite dimensions, showing that it outperforms i.i.d. test designs as
well as other recovery algorithms based on convex programming. |
2408.00386v1 | ## What comes after transformers? - A selective survey connecting ideas in deep learning ⋆
Johannes Schneider 1
University of Liechtenstein, Vaduz, Liechtenstein [email protected]
Abstract. Transformers have become the de-facto standard model in artificial intelligence since 2017 despite numerous shortcomings ranging from energy inefficiency to hallucinations. Research has made a lot of progress in improving elements of transformers, and, more generally, deep learning manifesting in many proposals for architectures, layers, optimization objectives, and optimization techniques. For researchers it is difficult to keep track of such developments on a broader level. We provide a comprehensive overview of the many important, recent works in these areas to those who already have a basic understanding of deep learning. Our focus differs from other works, as we target specifically novel, alternative potentially disruptive approaches to transformers as well as successful ideas of recent deep learning. We hope that such a holistic and unified treatment of influential, recent works and novel ideas helps researchers to form new connections between diverse areas of deep learning. We identify and discuss multiple patterns that summarize the key strategies for successful innovations over the last decade as well as works that can be seen as rising stars. Especially, we discuss attempts on how to improve on transformers covering (partially) proven methods such as state space models but also including far-out ideas in deep learning that seem promising despite not achieving state-of-the-art results. We also cover a discussion on recent state-of-the-art models such as OpenAI's GPT series and Meta's LLama models and, Google's Gemini model family.
Keywords: transformers, attention, state-space models, capsule networks, survey, review, deep learning, architectures, layers
## 1 Introduction
Transformers are widely regarded as the driving force behind artificial intelligence, e.g., of so-called foundation models[101] such as OpenAI's ChatGPT or early models such as BERT. Transformers are at the top of most machine learning benchmark leaderboard including computer vision, speech, and natural language processing. A major advantage of deep learning is its layered, modular structure. It enables construction of models from individual components in a flexible manner. Researchers have created a large selection
⋆ This is an extended version of the published paper by Johannes Schneider and Michalis Vlachos titled 'A survey of deep learning: From activations to transformers' which appeared at the International Conference on Agents and Artificial Intelligence(ICAART) in 2024. It was selected for post-publication.
of layers, architectures, and objectives. Keeping up with the rapid developments in AI models is a difficult task.
There exist multiple reviews with a narrow focus such as large language models (e.g. [77]) and convolutional neural networks (e.g. [56]). Previous studies [2,109,19,3] with a wider focus have become dated and miss new developments such as transformers and self-supervised learning. Furthermore, no survey or novel text book such as [11] has focused on alternatives towards transformers. Multiple works have discussed particular shortcomings of transformers such as computational inefficiency, e.g., due to quadratic time complexity of (naive) self-attention [99,37,12]. Thus, there is no work that looks at recent deep learning holistically integrating general ideas in deep learning and transformers, emphasizing alternatives or ways ahead. However, taking a broader and more holistic look at different (sub-)disciplines can be highly beneficial: For instance, NLP and computer vision have reciprocally shaped one another; CNNs first emerged in computer vision and subsequently found applications in NLP, whereas transformers originated in NLP and were later incorporated into computer vision. Consequently, breaking down the barriers between disciplines proves to be highly advantageous.
This paper is driven by the goal of examining the recent advancements in deep learning from an more encompassing perspective, rather than concentrating on a specific specialized field. We contend that such an approach is crucial, especially as significant new developments have decelerated; currently, the vast majority of models employs the 'old' but still mostly state-of-the art transformer architecture presented already in 2017[123].
Providing a comprehensive overview of the field is challenging, if not impossible, given the vast number of articles published annually and the constant expansion of pertinent topics. Our strategy is to choose influential works through (i) usage statistics, (ii) specialized surveys, and (iii) for transformer alternatives, we rely on public debates (as well as our own literature search). We also offer an invigorating discussion of shared design patterns across areas that have been successful. 1
## 2 Overview
Figure 1 outlines the topics addressed in this survey. In Part 1, We start our discussion with the key topic of this survey, i.e., transformers. We refer readers with lack of knowledge of basic architectural components of transformers such as attention to Part 2 of this work or text books[11,29]. However, Part 2, is also recommended for more experienced readers to reflect on recent ideas in deep learning. In Part 1, we elaborate on current architectures including commercial and open-source models but also go beyond by discussing more far-out ideas in deep learning. The implementations of these ideas have often not yet shown state-of-the-art results or even near state-of-the-art results, but nevertheless provide interesting ideas that might just need further thoughts.
In Part 2, our exploration focuses on deep learning design, encompassing both objectives and training methods. Particular emphasis has been placed on work validated by
1 This paper extends the conference paper [103] by engaging more deeply in discussing transfromer alternatives among other aspects
Fig. 1: Overview: Part 1 deals with transformers, Part 2 with deep learning (Lower graph based on [103])
usage metrics from the well-known platform "Paperswithcode.com." Our inclusion criteria feature impactful research published since 2016 and emerging contributions from 2020 onward that have quickly risen to prominence.
The coverage of each subject area is influenced by the volume of recent research and its fundamental importance. We omit discussions on data or computational strategies, such as data augmentation, model compression, and distributed machine learning, due to space constraints. This necessitated a discerning selection of model types, excluding significant categories like multi-modal models and autoencoders.
## 3 Transformers
Transformers have emerged as the leading architecture in deep learning. When paired with self-supervised training on extensive datasets, they have achieved top performance across numerous benchmarks in NLP (refer to [67] for a detailed survey) and computer vision (as outlined in [39,57]). Introduced in 2017 [123], numerous adaptations have been developed to address challenges such as computational demands and data efficiency in the original transformer model. We discuss classical and current models in Section 3.1, emerging alternatives to transformers and trends in Section 3.2 and more far-out ideas going beyond transformers in Section 3.3.
It is often stated that transformers possess lower inductive bias (e.g., compared to CNNs and RNNs), making them more adaptable. Consequently, they need a larger volume of training data to offset the reduced inductive bias. Transformers are commonly pre-trained using self-supervised learning, and oftentimes fine-tuned towards specific tasks on labeled data. Classical deep learning sequence models such as recurrent neural networks (RNNs) often exhibit a time-dependency upon inference and training, e.g., after processing of one token a hidden state must be updated before the next token can be processed. This limits parallelization and leads to other issues such as vanishing gradients for long sequences. Transformers, in contrast, allow to process the entire sequence at once using self-attention to predict the next token, i.e., they are much more parallelizable than sequential models such as RNNs. Unfortunately, naive implementation of self-attention require quadratic run-time, which makes them difficult to use for long sequences. Ideas like mixture of experts (MoE) transformer-based model have been employed to reduce computational burdens (especially during inference). MoE uses a learned routing function to select only a subset of the model for processing, which saves on computation. The idea of MoE dates back well before the year 2000 [134]. It has gained traction more recently as a means to reduce computational effort, in particular, during inference.
## 3.1 Classical and state-of-the-art transformer architectures
The original transformer[123], designed for NLP, utilizes an encoder and decoder similar to earlier recurrent neural networks. The architecture features multiple layers of transformer blocks, as depicted in Figure 2. Essential components include multi-head attention, layer normalization, and skip connections. For enhanced efficiency, several commercial and open-source models adopt the mixture of experts technique, which selectively computes outputs from a subset of models within an ensemble. Additionally, positional encodings and input embeddings are crucial. The absolute positional encodings PE for position pos in [123] are calculated using sinusoidal functions with varying frequencies:
$$P E ( p o s, 2 i ) = \sin ( p o s / 1 0 0 0 0 ^ { 2 i / d } )$$
$$P E ( p o s, 2 i + 1 ) = \cos ( p o s / 1 0 0 0 0 ^ { ( 2 i ) / d } )$$
where i represents the encoding dimension and d the total number of dimensions. This method was chosen because it allows relative positions, potentially as significant as absolute positions, to be a linear function of the absolute position encodings.
Bidirectional Encoder Representations from Transformers(BERT) [18]:The encoder of the transformer architecture generates contextual word embeddings through a masked language model pre-training objective and self-supervised learning. In this approach, the model is tasked with predicting randomly selected, masked input tokens based on their surrounding context, thus providing it with bidirectional information-that is, access to tokens both before and after the masked words. Unlike traditional next-word prediction methods, where no tokens following the predicted word are
Fig. 2: Transformer with the four basic blocks on top and the encoder and decoder at the bottom Figure from [103]
presented. Additionally, the model is required to determine whether a pair of sentences ( A B , ) are consecutive sentences from the same document or two unrelated sentences. Once pre-trained via self-supervised methods, the model can be fine-tuned for specific tasks using labeled data.
The original BERT model has undergone various improvements, for instance, [98] minimized BERT's computational demands, and [69] extended training durations, utilized longer sequences, larger batches, and more data. These enhancements have led to more robust and generalizable representations.
Text-to-Text Transfer Transformer (T5)[92] considers all text-based language models as systems that produce output text from an input text. This differs from BERT[18] through the use of causal masking in training, which blocks the network from seeing "future" tokens of the target. Additionally, T5 diverges in its pre-training tasks.
BART [63] is a denoising autoencoder used for pretraining sequence-to-sequence models with a conventional transformer-based machine translation architecture. It has proven effective in tasks such as language generation, translation, and comprehension. The training involves corrupting text using a variety of noising functions, including token deletion, masking, sentence permutation, and document rotation. The learning process involves reconstructing the original text from its corrupted form. The variety of noising techniques utilized can be attributed to BART's adaptation and generalization of concepts from previous models like BERT and GPT[89], combining a bi-directional encoder (similar to BERT) with an autoregressive decoder (similar to GPT).
XLNet [132] This approach merges the benefits of autoregressive modeling, as seen in GPT, where it predicts the next token, with the denoising auto-encoding capabilities of BERT[18], which involves reconstructing x from a noisy input ˆ created by maskx ing words in x . It achieves this by using a permutation language model that samples a permutation of Z = z 0 , z 1 , ..., zT -1 of the sequence ( 0 1 2 , , , ..., T -1 ) aiming for the objective:
$$\max p ( u _ { z _ { T } } | u _ { z _ { 0 } }, \dots, u _ { z _ { T - 1 } } ) & & ( 3 )$$
Importantly, the inputs themselves are not actually permuted, which would be impractical and inconsistent with later fine-tuning tasks. Instead, the permutation influences the attention mask to ensure that the factorization order determined by Z is preserved.
The Vision Transformer [21] splits an images into small patches, each of which is flattened and embedded linearly along with position embeddings. These embedded vectors for each patch are then processed by a standard transformer encoder.
The Swin Transformer [70] constructs hierarchical feature maps in computer vision applications instead of just a single resolution feature map. Additionally, it limits the computation of self-attention to within a local window, which helps reduce the overall computation time.
PaLM 2: The original PaLM model[17], containing 540 billion parameters, parallels other prominent models like GPT-3 in scale. Its technical advancements primarily focus on the scalability of model training, enabling a single model to be trained efficiently across tens of thousands of accelerator chips. Slight modifications were made to the original transformer architecture[123], such as incorporating SwiGLU activations, represented by
$$S w i s h ( x W ) \cdot x V$$
, where Swish is defined in Eq. 22. Additional changes include alternative positional embeddings optimized for longer sequences, multi-query attention for faster computation, the elimination of biases for improved training stability, and shared input-output embeddings.
PaLM 2[30], the more advanced successor of PaLM, distinguishes itself by utilizing a different mixture of datasets that incorporate a broader range of languages and domains, such as programming languages and mathematics. Although it uses the traditional transformer architecture, it operates with a smaller model size yet employs more computational resources for training. Moreover, it broadens its pre-training objectives beyond simple next-word or masked-word prediction, incorporating a variety of tasks to enhance its performance.
OpenAI's GPT to GPT-3 on to ChatGPT 3.5 and GPT-4o : The first version of Generative Pre-trained Transformer (GPT) is based only on the transformer's decoder to predict tokens one after the other. GPT[89] initially undergoes unsupervised pre-training, typically followed by supervised fine-tuning. The pre-training process involves a large corpus of tokens U = ( u 0 , u 1 , ..., un -1 ) and focuses on maximizing the probability of predicting subsequent tokens based on prior ones:
$$L ( U ) = \sum _ { i } \log p ( u _ { i } | u _ { i - k }, \dots, u _ { i - 1 } )$$
where k represents the context window size, and the conditional probability is computed using a neural network, specifically a multi-layer transformer decoder[68] that omits the encoder as per [123]. Additionally, this method reduces the memory usage of the attention mechanism.
Building on the foundational GPT model, GPT-2 [90] introduces a few key enhancements, such as relocating layer normalization to the beginning of each sub-block and adding an additional normalization after the last self-attention block. Moreover, the initialization of residual weights has been adjusted, and there has been an increase in the vocabulary, context window, and batch sizes.
GPT-3[13] maintains a structure nearly identical to GPT-2, but it expands dramatically in scale, with more than 100 times the number of parameters and variations in the volume of training data.
(Chat)GPT-3.5 [82] is closely related to InstructGPT[85], designed with a focus on adhering to user directives. InstructGPT refines GPT-3's capabilities through a dualphase fine-tuning approach: (i) guided by demonstrations from labelers using supervised learning and (ii) leveraging human evaluations of model responses in a reinforcement learning framework. ChatGPT adheres to this methodology, where (i) supervised learning involves human AI trainers engaging in scripted dialogues, where humans act as both the user and the AI agent. These dialogues are subsequently merged with the InstructGPT conversations, which have been reformatted into dialogues. In Phase (ii), AI trainers assess the quality of responses generated by ChatGPT, choosing from multiple possibilities for a randomly chosen model-generated text. This ranking informs the reinforcement learning phase to refine the model's performance.
The technical specifications of GPT-4, the follow-up to ChatGPT-3.5, have not been officially released as per the technical report[83], although some unofficial sources [73] claim to have insights. The report suggests that GPT-4 retains many characteristics of ChatGPT 3.5 but extends its capabilities to multi-modal functions, allowing it to process images as well. It highlights significant enhancements in training efficiency and a notable achievement in predicting the performance of large-scale models based on the outcomes of smaller models, which may have been trained on less data. This aspect is particularly crucial given the significant impact of computational costs and time on the development of extensive deep learning models.
Later models by OpenAI such as GPT-4 Turbo and GPT-4o [84] focused on improving efficiency and response times and multi-modality, including audio, with comparable performance or modest gains on benchmarks. While there are no technical details available, models were likely down-sized and the amount of training was increased, possibly with new or improved data.
Meta's Open source LLama series: LLama is Meta's open-source model family of LLMs, ranging from LLama[120], LLama-2[121] on to LLama-3.1[76]. The LLama3 paper also discussed multi-modal experiments. The largest model LLama 3.1 405B achieves similar performance as commercial models such as GPT4o and Claude Sonnet 3.5[76]. All models are dense transformers rather than MoE to improve training stability. Improvements, in particular for LLama-3 originated from scaling data, models and training and changes in post-processing, e.g, using direct preference optimization (DPO)[91] rather than reinforcement learning. That is, LLama-3 included no major ar-
chitectural changes compared to prior LLama models and, in turn, to the original transformer architecture or self-supervised pre-training proceducre. Minor updates were the usage of grouped query attention [1] and preventing self-attention between different documents (using an attention mask). Architectural details including hyperparameters are provided in [76].
Gemini, Claude and other models: The Gemini family has been briefly described in a technical report [117] and the version 1.5 family in [95]. However, technical details are mostly missing. It is also based on the transformer architecture by Vaswani et al. with improvements in architecture such as multi-query attention[107] and optimization to help training at scale and improve efficiency. It is multi-modal including video. The version 1.5 improved long-context understanding of inputs and it introduced a sparse mixture of expert (MoE) transformer-based model. The company 'Mistral.AI' also launched Mixtral, an opensource MoE model[50]. GPT-4 is also alleged to be based on MoE with eight experts [73]. Though Anthropic's Claude models are claimed to perform well, there was no official information on its architecture or training but only a model card [4].
## 3.2 Trends and Alternatives to Transformers
The original transformer architecture is still the key model for LLMs and a key driver for generative AI - as elaborated in the prior section 3.1. The prior section also highlighted a few trends that emerged in the last few years to improve transformers. Many focus on computational aspects by altering existing transformers, e.g., using a MoE to reduce computation or altering attention mechanisms, while others are not driven by computation in particular and suggest more radical ideas, e.g, by combining older models like RNNs with new elements of transformers.
Combining Sequence Models and Transformers : That is, there is a stream of research that aims to take elements from sequence models such as RNNs and LSTMs to transformers. However, one alleged success factor for transformers in the first place is their parallelizability in contrast to classical sequence models.
[8] proposed xLSTM, which improves on the classical LSTM cell by using exponential gating with memory mixing and a new memory structure. It allows for improved parallel training. While the results are promising, they stem mostly from smaller models (1.3B) and, thus, need to be validated using larger models.
Receptance weighted key value (RWKV) [87] combines the well parallelizable training of transformers with the fast inference of RNNs. For inference they achieve linear time complexity in sequence length S and memory needs linear in the (embedding) dimension. In contrast, classical transformer require a factor S more for both time and space(memory). The architecture bares similarity with an ordinary transformer but they use a variant of linear attention to focus on channels rather than the classical dotproduct token attention.
Retentive networks[115] replace multi-head attention with multi-scale retention. They derive it from a recurrent model by adding context awareness. In a recurrent setting, a state s is updated using a matrix Q . Context awareness means that the matrix Q
is also mulitplied by the context X derived from the current input. They also perform matrix diagonalization to obtain an easily parallizable formulation.
## State-space models - Replacing attention with a merger of CNNs and RNNs :
State-space models (SSM) are well known in natural science dating back more 60 years[51]. They can be written as x ′ ( ) = t Ax t ( ) + Bu t ( ) , y t ( ) = Cx t ( ) + Du t ( ) with matrixes A B C , , , and D . They have been adopted for sequence modeling[35], where they can be interpreted as 'a combination of recurrent neural networks (RNNs) and convolutional neural networks (CNNs), with inspiration from classical state space models'[33]. In particular, [34] focuses on long range dependencies leading to the so-called structured state space sequence model (S4) [34]. Technically, they apply a low-rank correction to matrix A allowing the matrix to be diagonalized stably. Special SSM layers have also been proposed for LLMs achieving comparable performance on small models [27]. Amajor evolution is MAMBA[33], which neither requires attention nor MLPs. It introduces a selection mechanism that allows to select data in an input-dependent manner. They also create a hardware aware algorithm and simplify prior SSM architectures by merging with the MLP layer into a single block.
Forgoing attention based on MLPs only and gated convolutions : One of the simplest architectures relies on multi-layer perceptrons (MLP)s only. The MLP-mixer[118] has been shown to achieve state-of-the-art results in computer vision without convolutions and self-attention. It relies on MLPs that are applied on spatial locations or feature channels instead. It also maintaining other common architectural elements such as skip-connections and layer normalization. Recent work has also improved on MLPs for mixing, e.g., by using Monarch matrices [26].
A drop-in replacement for attention is Hyena [88]. It is based on long convolutions and data-controlled gating. The paper also discusses three factors that must be fulfilled by attention alternatives but are often missed, i.e., (i) data control: Attention is a linear data controlled operator, (ii) sublinear parameter scaling: parameters (of attention layers) does not depend on sequence length, and (iii) unrestricted context: interaction between any tokens is possible, e.g., there is no locality.
Models to systems - Integrating planning and external tools : While a single LLM is often seen or believed to be capable of doing any task, there are a number of approaches that aim at building systems rather than improving only the deep learning model in isolation. That is, LLMs are enhanced with other components that interact to improve its performance. For example, [52] claim that large language models (LLMs) cannot plan but can generate ideas that are critiqued by separate models, which may be domain-specific and not LLMs. Ideas like Retrieval Augmented Generation (RAG)[64] have altered the way how LLMs process prompts by first enhancing them with external knowledge, which helped to reduce hallucinations. The usage of external tools by the LLMsuch as calculators has also improved the performance of LLM-based systems for many tasks[100].
## 3.3 Far-out models and ideas
The prior section contained some interesting ideas that have arleady shown promising outcomes compared to transformers in empirical evaluations. However, one might also
seek to leverage more revolutionary ideas that have been proposed in the context of deep learning but have not yet yielded the desired success.
Capsule networks - Returning vectors instead of scalars : Capsule networks have been proposed in 2011[45]. The essential idea is that a neuron outputs a vector rather than a scalar. Hinton et al. motivate the usage of a vector output in the context of computer vision. They argue that a (visual) feature is often characterized by multiple instantation parameters such as position, scale, lighting, etc. that resemble a vector rather than just a scalar. A capsule learns an implicitly defined visual entity across some limited transformations such as lighting or scaling. It returns the probability that the entity is present and a set of instantation parameters (e.g., lighting or scaling). They argue that a key advantage of this approach is the recognition of whole-part relationships, e.g., to detect a face one might identify a mouth and nose and both must have appropriate scaling and orientation. The training of the network requires pairs of transformed images to learn capsules that can extract pose parameters from inputs. Capsules themselves might consist of classical elements such as dense layers and gates. Capsule networks have been further improved since their introcduction in 2011[96,46]. They managed to achieve good performance on simple daasets such as MNIST but failed to deliver close to state-of-the-art performance on challenging datasets such as ImageNet.
Spiking neural networks - Adding temporality through non-synchronous activating neurons : Spiking neural networks [72] exist for more than 25 years, achieving remarkable, but no state-of-the-art results. Spiking neural networks add a temporal dimension to neural networks. A neuron activates once its 'potential' exceeds a threshold through transmission of spiking signals to its adjacent neurons. The spike increases their potential. Potentials decrease upon sending spikes but also spontaneously over time (without transmission). Thus, in contrast to non-spiking networks, transmission of signals in the network can occur in non-periodic, non-synchronized cycles among neurons. As such the time, when a signal occurs also resembles information in addition to the actual information in the spike. This effectively allows to perform time-coding as common distributed systems [104].
Neurosymbolic AI - Leveraging domain specific (symbolic) languages : There is also an (ongoing) debate on the value of symbolic approaches in conjunction with neural networks relying on continuous representations. For example, in 2017 a neuro symbolic approach for program synthesis was proposed[86] that can generate program code in a domain specific language from input-output examples. While the synthesis appears trivial compared to current program skills of LLMs, the paper still contains a number of interesting ideas, e.g., related to how to search the space of possible programs and how to incrementally refine programs. Other neuro-symbolic approaches have also performed concept extraction based on neural networks and used domain specific languages to perform quasi-symbolic program execution [74]. A comprehensive overview with a focus on NLP is given in [38].
Embrassing ideas from the mathematics community : Classical ideas from mathematics have also entered the domain of deep learning. For example, residual neural networks (ResNets) [42] essentially solve ordinary differential equations(ODE) [14]. A simple ODE is dy dt = f ( y t , ) and the inital condition y t ( 0 ) = y 0. Its solution yn at dis-
cretized time steps can be computed (recursively) using the Euler method:
$$y _ { n + 1 } = y _ { n } + f ( y _ { n }, t _ { n } ) \cdot ( t _ { n + 1 } - t _ { n } )$$
A layer n in a residual network essentially performs the same computation: yn + 1 = yn + ( f yn , t n ) Thus, a sequence of residual blocks can be interpreted as a solution of the ODE with the Euler method with the initial condition y ( 0 ) = X , where X is the input to the network. Viewing ResNets as ODEs allows to utilize mathematical tools such as ODE solvers and, in turn, might also have benefits with respect to memory and parameter efficiency. The idea of reformulating learning systems has also led to performance improvements [40]. That is, e.g., [40] introduce time-continuous RNNs described as a linear first order system.
Embrassing ideas from neuroscience and psychology : While AI does not seek to rebuild the brain, (articial) neural networks structurally still share similarity with the real neural networks in our heads. Thus, almost any finding in the neuroscience community might be used as an inspiration for improving artificial neural networks. The main hinderness to adopting neuroscience in deep learning research is 'that we simply do not have enough information about the brain to use it as a guide'[29]. For instance, the bain has (neurological) feedback loops: A neuron receives and processes a signal, and can then send a response back to the original source. While numerous attempts have been made to leverage this idea, overall success has been limited, i.e., it is essentially a call for more research. For example, [53] argued that feedback (or recurrent circuits) are critical for fast object identification. In another idea from psychology, a reflective architecture was introduced that leverages explanations computed from gradients (GRADCAM) during learning but also during inference[102] drawing inspirations from Kahneman's fast thinking (System 1) and slow (reflective) thinking (System 2). The idea of ordering samples by difficulty to improve learning of deep learning models, i.e., curriculuum learning [10], can be traced back to the idea of shaping in animal training (e.g., [111]).
## 4 Loss functions and Optimization
We elaborate on fundamental loss functions and optimizers.
## 4.1 Loss Functions
Loss functions, as reviewed in [126], typically include several terms including a regularization term. While these functions are generally tailored to specific tasks, certain universal principles can be applied across various tasks. It is common to combine multiple loss terms using a weighted approach. Frequently, enhancements to previous studies are achieved merely by modifying the loss function.
The Triplet Loss [20] was introduced for Siamese networkss, though its roots trace back to earlier work[105]. The overarching principle involves comparing a given input with both a positive and a negative input, aiming to maximize the association with
positively related inputs and minimize it with negative ones. It operates on input pairs ( x y , ) , each processed by a distinct but architecturally identical network. The goal is to maximize the joint probability p x y ( , ) for all pairs ( x y , ) :
$$L ( \mathcal { V } _ { p }, \mathcal { V } _ { n } ) = - \frac { 1 } { | \mathcal { V } _ { p } | \cdot | \mathcal { V } _ { n } | } \sum _ { x \in \mathcal { V } _ { p } y \in \mathcal { V } _ { n } } \log p ( x, y )$$
$$= - \frac { 1 } { | \mathcal { V } _ { p } | \cdot | \mathcal { V } _ { n } | } \sum _ { x \in \mathcal { V } _ { p } y \in \mathcal { V } _ { n } } \sum _ { y \in \mathcal { V } _ { n } } \log ( 1 + e ^ { x - y } )$$
Here, V p and V n are the positive and negative score set respectively.
The Supervised Contrastive Loss [59] groups clusters of points from the same class together in the embedding space while pushing apart samples from different classes. It is designed to utilize label information more effectively than cross-entropy loss, optimizing the separation and aggregation based on class identity.
$$\mathcal { L } _ { i } ^ { s u p } = \frac { - 1 } { 2 N _ { \tilde { y } _ { i } } - 1 } \cdot$$
̸
̸
$$\sum _ { j = 1 } ^ { 2 N } \mathbf 1 _ { i \neq j } \cdot \mathbf 1 _ { \tilde { y } _ { i } = \tilde { y } _ { j } } \cdot \log \frac { \exp ( z _ { i } \cdot z _ { j } / \tau ) } { \sum _ { k = 1 } ^ { 2 N } \mathbf 1 _ { i \neq k } \cdot \exp ( z _ { i } \cdot z _ { k } / \tau ) }$$
where Nyi ˜ denotes the total number of images in the minibatch with the same label ˜ y i as the anchor i . The overall loss is calculated as the sum of the losses for all anchor points i , i.e., L = ∑ i L sup i . This loss formulation is particularly advantageous for supervised learning due to its:
- -Ability to generalize across an arbitrary number of positive examples.
- -Increased contrastive effectiveness as the number of negative examples grows.
The Cycle Consistency Loss [136] is specifically designed for unpaired image-to-image translation using generative adversarial networks. It facilitates the learning of mappings between two distinct image domains X and Y . Optimizing the loss supports the learning of mappings G : X → Y and F : Y → X so that one reverses the other, i.e., F G x ( ( )) ≈ x and G F y ( ( )) ≈ y .
$$L ( G, F ) = \mathbb { E } _ { x \sim p _ { d a t a } ( x ) } [ | | F ( G ( x ) ) - x | | _ { 1 } ]$$
$$+ \mathbb { E } _ { y \sim p _ { d a t a } ( y ) } [ | | G ( F ( y ) ) - y | | _ { 1 } ]$$
Focal Loss [65] modifies the standard cross-entropy loss to concentrate learning efforts on hard-to-classify samples. It incorporates a factor ( 1 -p ) γ , where p is the probability of a sample from the cross entropy loss and γ is a freely adjustable parameter.
$$L ( p ) = ( 1 - p ) ^ { \gamma } \log ( p )$$
## 4.2 Regularization
Regularization techniques, as reviewed in [80], are critically beneficial for deep learning applications. Explicit regularization incorporates an additional loss term R f ( ) for the network f into the loss function L x ( ) for data ( xi , yi ) with a balancing parameter λ .
$$\min _ { f } \sum _ { i } L ( x _ { i }, y _ { i } ) + \lambda R ( f )$$
Implicit regularization encompasses all other forms of regularization, such as early stopping or employing a robust loss function. Common methods like L 2-regularization and dropout[113], which involves setting the activations of a random subset of neurons to zero, are among the most extensively used techniques.
Entropy Regularization [79] is designed to promote diversity, particularly through asynchronous methods in deep reinforcement learning [127,79]. This approach enhances the diversity of actions in reinforcement learning by avoiding excessive optimization towards only a small portion of the environment. Entropy is calculated based on the probability distribution of actions as determined by the policy as:
$$H ( x ) = \sum _ { x } \pi ( x ) \cdot \log ( \pi ( x ) )$$
Path Length Regularization [54] [54] focuses on enhancing the image quality for generative adversarial networks by ensuring consistent step lengths in the latent space and changes in the generated images. That is, changes in the latent space representation lead to (proportional) changes in the generated image (and vice versa). The principle behind this is to facilitate a fixed-size step in the latent space W to yield a consistent, non-zero change in image magnitude. This approach aims to improve the conditioning of GANs, which can simplify the process of architecture search and generator inversion. Gradients, with respect to w ∈ W , originating from random directions in the image space should be almost equal in length independent of w or the image space direction. The local metric scaling characteristics of the generator g : W → Y are stated by the Jacobian matrix Jw = δ g ( w ) / δ w . The formulation for the regularizer is:
$$\mathbb { E } _ { w, y \sim \mathcal { A } ( 0, \mathbf I ) } ( | | \mathbf J _ { w } ^ { \mathbf T } \mathbf y | | _ { 2 } - a ) ^ { 2 }$$
where y corresponds to random images with pixel values that follow a normal distribution, and w ∼ f ( ) z , where z is also normally distributed. The constant a is defined as the exponential moving average of || J T w y || 2. The paper also avoids the computationally expensive, explicit computation of the Jacobian.
DropBlock [28] selectively drops correlated regions within feature maps instead of dropping features independently. This method is particularly effective for convolutional neural networks, where feature maps tend to show spatial correlations and a real-world feature typically corresponds to a contiguous spatial area within these maps.
R 1 Regularization [75] penalizes the discriminator in generative adversarial networks based on the gradient to stabilize training:
$$R _ { 1 } ( \Psi ) = \frac { \gamma } { 2 } E _ { p _ { D } ( x ) } [ | | \nabla D _ { \Psi } ( x ) | | ^ { 2 } ]$$
Technically, the regularization term imposes penalties on gradients that are orthogonal to the data manifold, aiming to refine the learning process.
## 4.3 Optimization
Optimization, as explored in [114], involves determining the best set of network parameters to minimize the loss function. The two most widely recognized techniques are stochastic gradient descent (SGD) and Adam. Neither technique consistently surpasses the other across all scenarios in terms of generalization performance. SGD has historical roots extending back to the 1950s [60], whereas Adam was developed more recently, in 2014 [61].
Adafactor [108] enhances the efficiency of the Adam optimization algorithm by reducing its memory requirements. This is achieved by keeping only row- and column-wise statistics of parameter matrices, instead of storing detailed per-element information.
Layerwise adaptive large batch optimization (LAMB) [133] builds on Adam and accelerates training by utilizing large mini-batches. It applies per-dimension and layerwise normalization to further enhance the optimization process.
Decoupled Weight Decay Regularization for ADAM : AdamW[71] leverage a seemingly trivial observation: The original Adam optimizer updates weights with (L2-)regularization after calculating the gradients for Adam. However, it intuitively makes more sense that moving averages of gradients should exclude regularization.
RAdam and AMSGrad : Both techniques address the convergence issues associated with Adam. Rectified Adam (RAdam)[66] adjusts the variance of the adaptive learning rate, which is initially large. This approach is akin to the warm-up heuristic, where small initial learning rates are beneficial. AMSGrad[94] uses the maximum of past squared gradients instead of the exponential average to enhance stability and convergence.
Stochastic Weight Averaging : Naive averaging of weights from different epochs during stochastic gradient descent, using either a constant or cycling learning rate, has been shown to improve performance[49].
Two Time-scale Update Rule(TTUR) : TTUR[44] enhances the training of generative adversarial networks by employing separate learning rates for the discriminator and the generator. When the generator is fixed, the discriminator can reach a local minimum. This approach remains effective even if the generator converges slowly, for instance, by using a smaller learning rate. By allowing the generator to more deeply assimilate feedback from the discriminator before exploring new regions, this method aids in the convergence of GANs and can lead to improved performance. Besides ensuring convergence, the performance may also improve since the discriminator must first learn new patterns before they are transferred to the generator. In contrast, a generator which is overly fast, drives the discriminator steadily into new regions without capturing its gathered information.
Sharpness-Aware Minimization [25] minimizes both the loss value and sharpness, which enhances generalization. It seeks parameters within neighborhoods of low loss values, rather than focusing solely on parameters that individually exhibit low loss. The
loss function is defined as:
$$\min _ { w } \max _ { | | \epsilon | | _ { p } \leq \rho } L ( w + \epsilon )$$
## 5 Self, Semi-supervised and Contrastive learning
Semi-supervised learning utilizes a large volume of unlabeled data alongside a small amount of labeled data, as reviewed in [131]. Self-supervised learning, on the other hand, generates (pseudo) labels from artificial tasks, reducing the need for manually labeled data. Both approaches alleviate the burden of collecting human-annotated data. Combining self-supervised (pre-)training with fine-tuning on a small human-labeled dataset can achieve state-of-the-art results. This paradigm has significantly expanded in recent years (surveyed in [24]). It is often integrated with contrastive learning, which aims to learn the distinction between similar and dissimilar data. By automatically distorting data to varying degrees, creating "pseudo-labeled" data for self-supervised learning becomes straightforward.
The simple framework for contrastive learning (SimCLR)[15] maximizes the agreement between two inputs derived from different augmentations of the same data sample. Augmentations can include random cropping, color distortions, and Gaussian blur. A standard ResNet[42] is used to obtain representation vectors, which are then further processed through a simple MLP before applying the contrastive loss.
Bootstrap Your Own Latent (BYOL) [31] employs both an online network and a target network, each with identical architectures comprising an encoder, a projector, and a predictor, but with separate weights. The parameters of the target network are updated as an exponential moving average of the online network's parameters. The task of the online network is to predict the representation of the target network given an augmentation of the same input.
Barlow Twins [135] rely on an objective function that aims to reduce cross-correlation C between outputs for a set of image Y A and their distorted versions Y B as close to the identity as possible, i.e., the loss (including λ as a tuning parameter) is:
use an objective function aiming to minimize the cross-correlation C between outputs for a set of images Y A and their distorted versions Y B , bringing it as close to the identity matrix as possible. The loss function, incorporating λ as a tuning parameter, is given by:
̸
$$L = \sum _ { i } ( 1 - C _ { i, i } ) ^ { 2 } + \lambda \cdot \sum _ { i } \sum _ { j \neq i } C _ { i, j } ^ { 2 }$$
Momentum Contrast (MoCo) [41] constructs a dynamic dictionary via an encoder using unsupervised contrastive learning. During training, it performs look-ups, ensuring that an encoded query closely matches its corresponding encoded key while being dissimilar to other keys. The dictionary functions as a queue of data samples: for each mini-batch, new encoded samples are added, and the oldest mini-batch is dequeued.
The key encoder's parameters are updated using a momentum-based moving average of the query encoder, promoting consistency over time.
Noisy Student : [129] outlines a method where an CNN-based EfficientNet model is initially trained on labeled data. This trained model then acts as a teacher, generating pseudo labels for unlabeled images. Subsequently, a larger model is trained on the combined labeled and pseudo-labeled data. This process is iteratively repeated, with each student model becoming the teacher for the next iteration. During the training of the student model, techniques like dropout and data augmentation are employed to introduce noise, making the learning process more challenging and allowing the student to surpass the performance of the teacher.
FixMatch [112] predicts the label of a weakly-augmented image. If the confidence in this prediction exceeds a certain threshold, the model is then trained to produce the same label for a strongly-augmented version of the image.
## 6 Architectures and Layers
We elaborate on four key types of layers: activation layers, skip connections, normalization layers, and attention layers. This is followed by a discussion of various contemporary architectures based on transformers and graph neural networks.
## 6.1 Activation
Activation functions are typically non-linear and significantly influence gradient flow and learning. Early activation functions, such as sigmoid and tanh, were commonly used from the 1960s through the early 2000s. However, these functions can cause training difficulties in deep networks due to the vanishing gradient problem when they saturate. The introduction of the rectified linear unit (ReLU) in 2010[81] was a breakthrough. Although the original ReLU remains widely used, transformer architectures have introduced other activation functions and ReLU variants. Most of these alternatives share the qualitative behavior of ReLU, where outputs for negative inputs are of small magnitude and outputs for positive inputs are unbounded (see [5] for a survey).
Gaussian Error Linear Units (GELU) [43] These functions weigh inputs by their percentile (ReLUs only use the sign). The activation function is the product of the input and the standard Gaussian cumulative distribution function Φ ( x ) , i.e.,
$$\ G E L U ( x ) = x \cdot \Phi ( x )$$
The Mish activation[78] originates from systematic search inspired by Swish and ReLU:
$$f ( x ) = x \cdot \tanh ( s o f t ^ { + } ( x ) )$$
$$\text{with sof} t ^ { + } ( x ) \coloneqq \ln ( 1 + e ^ { x } )$$
In contrast, the Swish activation[93] is:
$$f ( x ) = x \cdot s i g m o i d ( \beta x )$$
Here β is a learnable parameter.
## 6.2 Skip connections
Skip connections were introduced for residual networks[42]. In their simplest form, the output y for an input x of a block L of layers with a skip connection is y x ( ) = L x ( ) + x . The term "residual" was used in the original paper because the layer L is tasked with learning a residua L x ( ) = H x ( ) -x rather than the desired mapping H itself. Since then, skip connections have been utilizied in a number of variations.
ResNeXt Block [130]: This split-transform-merge approach for residual blocks involves evaluating a set of residual blocks concurrently and then aggregating their outputs back into one output. A Dense Block [47] receives inputs from all preceding layers with matching feature map sizes and connects to all such subsequent layers.
Inverted Residual Block [97]: By reversing the channel width sequence to a narrowwide-narrow order from the original wide-narrow-wide configuration[42], and using depthwise convolutions for the wide layer, parameters are reduced, and residual blocks execute more quickly. Furthermore, the activation function of the last layer within a block is skipped.
## 6.3 Normalization
With the advent of batch normalization[48], the concept of normalization has significantly enhanced training speed, stability, and generalization in neural networks including almost all architectures. However, its necessity is debated[106]; for certain applications, careful initialization and learning rate adjustments may render normalization partially redundant.
The principle behind normalization is to transform a value x to a normalized value ˜, by x subtracting the mean µ and scaling by the standard deviation σ , i.e., ˜ x = x -µ σ . Normalization approaches differ in the computation of µ and σ , e.g., µ and σ can be computed across different channels.
Layer Normalization : Normalization statistics are calculated using summed inputs[6] for a layer L with | L | neurons as:
$$\mu = \frac { 1 } { | L | } \sum _ { i = 0 } ^ { | L | - 1 } a _ { i } \quad \sigma = \sqrt { \frac { 1 } { | L | } \sum _ { i = 0 } ^ { | L | - 1 } ( a _ { i } - \mu ) ^ { 2 } }$$
Unlike batch normalization, this method does not limit the batch size and eliminates inter-batch dependencies, making it suitable even for batch sizes as small as one.
LayerScale [119] has been established for transformers as a per-channel multiplication of outputs of a residual block with a diagonal matrix:
$$x _ { l ^ { \prime } } = x _ { l } + d i a g ( \lambda _ { 1 }, \dots, \lambda _ { d } ) \cdot S A ( \eta ( x ) )$$
$$x _ { l + 1 } = x _ { l ^ { \prime } } + d i a g ( \lambda _ { 1 }, \dots, \lambda _ { d } ) \cdot F F N ( \eta ( x ) )$$
SA is the self-attention layer, FFN is the feed forward network, and η the layer-normalisation (see Figure 2).
Instance Normalization [122] calculates for a four-dimensional input, such as an image of height H , width W , channels C , and batch size T :
$$\mu _ { t, c } = \frac { 1 } { H W T } \sum _ { t < T, w < W, h < H } x _ { t, c, w, h }$$
$$\sigma _ { t, c } = \sqrt { \frac { 1 } { H W T } \sum _ { t < T, w < W, h < H } ( x _ { t, c, w, h } - \mu _ { t, c } ) ^ { 2 } }$$
This method can be applied, for example, to normalize image contrast. Various adaptations exist, including a version that scales according to weight norms[55].
## 6.4 Attention
Attention mechanisms, explored in [12,37], enable the learning of relevance scores for inputs, mimicking cognitive attention processes. This allows certain parts of the inputs to be highlighted as highly important, while others may be ignored as irrelevant. Often, the importance of a specific input is influenced by its context; for example, the significance of a word in a text document usually relies on the words surrounding it.
## Scaled Dot-Product Multi-Head Attention [123]:
The use of dot products, coupled with down-scaling, has been highly effective in calculating attention scores. Attention takes a query Q , a key K and a value V as inputs and outputs an attention score:
$$\text{Att} ( Q, K, V ) = \text{softmax} ( \frac { Q K ^ { T } } { \sqrt { d _ { k } } } ) \cdot V$$
Employing multiple, independent attention mechanisms in parallel enables the system to focus on different aspects of the input simultaneously. In the case of multi-head attention, matrices W are learned to facilitate this process. Formally:
$$\text{MultiHead} ( Q, K, \mathbf V ) = [ h _ { 0 }, \dots, h _ { n - 1 } ] \mathbf W _ { 0 }$$
$$\text{where head} \, h _ { i } = \text{Att} ( \text{QW} _ { i } ^ { Q }, \text{KW} _ { i } ^ { K }, \text{VW} _ { i } ^ { V } )$$
Factorized (Self-)Attention [16] decreases both the computational and memory demands compared to the original 'full' self-attention[123], which allows every element to attend to every other prior input. In factorized self-attention, attention is limited to a subset of input elements. Formally, an output matrix is generated using a matrix of input embeddings X and the connectivity pattern S = { S 1 , ..., Sn } , where Si is the set of indices of input vectors attended to by the i th output vector.
$$\text{FacAtt} ( X, S ) = ( A ( \mathbf x _ { i }, S _ { i } ) ) _ { i \in [ 1, n ] } & & ( 3 1 )$$
$$a ( { \mathbf x } _ { i }, S _ { i } ) = \text{softmax} ( \frac { ( W _ { q } { \mathbf x } _ { i } ) \overset { \cdot \, \cdot \, \cdot } { K } _ { S _ { i } } ^ { T } } { \sqrt { d } } ) \cdot V _ { S _ { i } }$$
$$K _ { S i } = ( W _ { k } \mathbf x _ { j } ) _ { j \in S _ { i } } \quad V _ { S _ { i } } = ( W _ { \nu } \mathbf x _ { j } ) _ { j \in S _ { i } }$$
̸
For full self-attention S F i : = { | j j = i } , i.e., all indexes to inputs prior to the i th input. In contrast, factorized self-attention has p separate attention heads, where the m th head is given by a subset A i ( m ) ⊂ S F i with Si = A i ( m ) . For strided self-attention:
$$A _ { i } ^ { ( 1 ) } = \{ t, t + 1, \dots i \} \text{ for } t = \max ( 0, i - l )$$
$$A _ { i } ^ { ( 2 ) } = \{ j \colon ( i - j ) \mod l = 0 \}$$
This pattern is adequate, when structure is similar to the stride-like images. For data without a periodic structure like text, fixed attention can be preferable:
$$A _ { i } ^ { ( 1 ) } = \{ j \colon \lfloor j / l \rfloor = \lfloor i / l \rfloor \}$$
$$A _ { i } ^ { ( 2 ) } = \{ j \colon j \mod l \in \{ t, t + 1, \dots l \} \}$$
where t = l -c with hyperparameter c . For instance, with a stride of 128 and c = 8, all future positions greater than 128 can attend to positions 120-128, all greater 256 to 248-256, and so on.
Multi- and Grouped Query Attention [107,1] Multi-query attention(MQA)[107] uses multiple query but only one key head and one value head, which differs from standard muli-head attention. This helps in lowering memory bandwidth for loading keys and values but it reduce quality and training stability. MQA can be added to trained models without MQA with (relatively) little compute [1]. Grouped query attention aims at striking a balance between a single and many heads. Per head they use multiple queries rather than just one. This allows to lower quality gaps.
A Residual Attention Network (RAN) [125] capitalizes on the concept of skip connections and is composed of two branches: a mask branch and a trunk branch. The trunk branch is responsible for processing features and can be any type of network, while the mask branch determines the weights for these features. The output of the attention module is
$$H _ { i, c } ( x ) = ( 1 + M _ { i, c } ( x ) ) \cdot F _ { i, c } ( X )$$
where i is a spatial position and c is a channel. M x ( ) should be approximatedly 0, H x ( ) approximates original features F x ( ) .
Large Kernel Attention [36] break down a (large-scale) convolution into three more manageable components: a depth-wise dilated convolution, a non-dilated depth-wise convolution, and a 1x1 convolution across channels. Subsequently, an attention map is derived from the outputs of these convolutions.
Sliding Window Attention [9] This approach focuses on enhancing the efficiency of attention mechanisms in terms of both time and memory by reducing the number of input pairs considered. Specifically, for a designated window size w , each token attends to w 2 tokens on either side.
## 6.5 Graph Neural Networks
Graph neural networks, reviwed in [128,58], can be viewed as an extension of CNNs and transformers to accommodate graph-structured data. These networks process information represented as nodes interconnected by edges. We delve into various graph
models, focusing on how to derive node embeddings that are applicable to downstream tasks.
Graph Convolutional Networks (GCN) [62] use CNNs for semi-supervised learning. They approximate spectral graph convolutions using polynomials of order k , which a CNN can compute with k linear layers. This also implies that dependencies are localized, i.e., only up to nodes of distance k from a node.
Scalable Feature Learning for Networks (Node2Vec)[32] is designed to learn feature vectors that effectively preserve the neighborhood characteristics of nodes within a graph. This method employs random walks to generate samples of neighborhoods, allowing it to capture and represent nodes based on their roles or the communities to which they belong.
Graph Attention Networks [124] utilizes masked self-attention layers that enable nodes to adaptively focus on the features of their neighboring nodes. Specifically, node j assigns importance scores to the features of node i based on their connectivity. The use of masking ensures that only connected node pairs are considered. Unlike GCNs, this method allows for different levels of importance to be assigned to nodes within the same neighborhood. Additionally, it avoids the expensive matrix operations associated with eigendecompositions.
TuckER [7] focuses on using factorization for link prediction within a knowledge graph, where knowledge is structured as (subject, relation, object) triplets. The objective is to predict the existence of a relationship between two entities. This approach models the graph as a binary tensor, with the dimensions representing subjects, relations, and objects. Tucker decompositions are then employed to break down this binary tensor into a core tensor and separate embedding matrices for subjects, relations, and objects.
Embedding by Relational Rotation (RotatE)[116] is used for predicting missing links in knowledge graphs, similar to the previously described TuckER[7], but with a focus on modeling additional relational properties like composition and inversion. In this method, entities are embedded in a complex space, and relations are treated as elementwise rotations. These rotations are optimized to effectively transition from one entity to another within the complex embedding space.
Graph Transformer [23] extends the original transformer architecture to handle graph structures by implementing attention mechanisms that consider the connectivity of each node's neighborhood. This adaptation generalizes the concept of position encoding to suit graph data. The model also replaces layer normalization with batch normalization and introduces the capability to learn representations for edges, in addition to node representations.
## 7 Discussion
Our findings suggest that despite many small and creative innovations since the original transformer architecture, there have not been any significant "breakthrough" discoveries that have led to much better leaderboard results. The improvements on models such as large language models within the last few years have been characterized by the enlargement of existing networks such as GPT, the increase of data volume (and quality),
focus on computational efficiency, and a shift towards self-supervised learning. This could indicate a need for more daring approaches to research rather than incremental improvements of existing works and it raises the question: Are transformers all we can do?
Combining different elements or pursuing more radical ideas as outlined in this work could be one way to go beyond transformers. But advancing novel ideas such as state-space models and ideas that have existed for a while such as capsule networks, but have not lived up to their hopes, is not an easy task. We have identified a few general patterns that have been proven effective in many areas and might help in accomplishing this task:
- -Multi-X refers to the strategy of deploying the same component repeatedly in parallel configurations, such as employing several residual blocks (ResNeXt) or utilizing multi-head attention mechanisms or multiple models as for mixture of experts. This concept is closely aligned with the principles of "ensemble learning."
- -Higher order layers involve more complex operations than the linear layers and simple ReLU functions typically used in classical CNNs and MLPs. Examples of these more advanced layers include Mish or attention layers, which facilitate deeper and more nuanced processing.
- -Data controlled gating is an instance of a higher order layer. The idea of data controlled gating, e.g., depending on some (possibly transformed) input X , another input X ′ is deemed relevant or ignored, seems powerful. This idea can be found in attention but also in LSTM layers and GELU, where we have 'self-gating'. On a larger scale also mixture of experts perform data controlled gating and, more generally, in transistors. Data controlled gating can also often be interpreted as Weighing functions , which involves the use of parameterized functions to weigh inputs. Instead of simply aggregating inputs, these functions assign weights to them, often derived from learned parameters. These functions act as "gates," allowing the flow of information only within a specific range of input parameters.
- -Moving average refers to the technique of averaging weights, as seen in methods like SGD and BYOL.
- -Decompose refers to breaking down matrices into simpler components, as exemplified by methods like TuckER and large kernel attention.
Furthermore, the transformer seems to violate the 'no free lunch theorem' by Wolpert saying that no model can be best for all (generative) tasks. While fine-tuning models on specific data can yield better (transformer) models for specific tasks, it is not clear, whether specialized architectures might substantially outperform the general transformer in certain sets of tasks, where it seems to outperform. This is analogous to activation functions, where no activation outperforms on all datasets and within all models [22].
Our survey focused on key design elements in building deep learning models. Taking a practical approach, we chose to ignore theoretical works, which should be further explored in future studies. For example, [110] derived a theoretical framework for dynamical systems allowing to better compare model classes such as linear attention and state space models.
Our survey intentionally focused on more recent, yet well-established works, which could be seen as either a strength or a limitation. The selection of papers for the deep
learning part was guided by a prominent platform that offers leaderboards. The rise of such platforms, which allow the upload of papers and models and provide information on citations and rankings, offers a fresh perspective compared to traditional survey methods that often choose papers more arbitrarily. Although this approach benefits readers looking for "what works well and what is very promising," it may overlook innovative ideas that need more research to fully demonstrate their potential. This could contribute to the "winner-takes-all" dynamic, reinforcing already successful ideas. Furthermore, such platforms might undergo a hype-cycle, e.g., become very popular for a while before vanishing. Nevertheless, given the vast number of papers, some form of selection is essential for conducting a comprehensive survey of deep learning.
We recognize that online platforms offering leaderboards and related features are highly valuable to the research community and should continue to be developed. However, we found that manual verification was necessary, such as double-checking relevance with Google Scholar citations and reviewing surveys and papers, to identify works and methods that were not accurately listed on the platform.
## 8 Conclusions
We have provided a concise yet thorough overview of the deep learning design landscape, with a focus on transformers and their potential successors. Key works from various significant areas that have emerged in recent years have been summarized. We believe that our holistic overview in a single paper can establish connections that may inspire novel ideas. Additionally, we have identified four patterns that characterize many improvements in this field. To further advance the development of artificial intelligence, it is crucial to generate fundamentally new and successful approaches, as recent improvements, though numerous and often very creative, have primarily been incremental.
## References
- 1. Ainslie, J., Lee-Thorp, J., de Jong, M., Zemlyanskiy, Y., Lebrón, F., Sanghai, S.: Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245 (2023)
- 2. Alom, M.Z., Taha, T.M., Yakopcic, C., Westberg, S., Sidike, P., Nasrin, M.S., Hasan, M., Van Essen, B.C., Awwal, A.A., Asari, V.K.: A state-of-the-art survey on deep learning theory and architectures. electronics (2019)
- 3. Alzubaidi, L., Zhang, J., Humaidi, A.J., Al-Dujaili, A., Duan, Y., Al-Shamma, O., Santamaría, J., Fadhel, M.A., Al-Amidie, M., Farhan, L.: Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions. Journal of big Data (2021)
- 4. Anthropic: The claude 3 model family: Opus, sonnet, haiku. Online (2023), https: //www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/ Model\_Card\_Claude\_3.pdf , accessed: 2024-07-20
- 5. Apicella, A., Donnarumma, F., Isgrò, F., Prevete, R.: A survey on modern trainable activation functions. Neural Networks (2021)
- 6. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv:1607.06450 (2016)
- 7. Balaževi´ c, I., Allen, C., Hospedales, T.M.: Tucker: Tensor factorization for knowledge graph completion. arXiv:1901.09590 (2019)
- 8. Beck, M., Pöppel, K., Spanring, M., Auer, A., Prudnikova, O., Kopp, M., Klambauer, G., Brandstetter, J., Hochreiter, S.: xlstm: Extended long short-term memory. arXiv preprint arXiv:2405.04517 (2024)
- 9. Beltagy, I., Peters, M.E., Cohan, A.: Longformer: The long-document transformer. arXiv:2004.05150 (2020)
- 10. Bengio, Y., Louradour, J., Collobert, R., Weston, J.: Curriculum learning. In: Proceedings of the 26th annual international conference on machine learning. pp. 41-48 (2009)
- 11. Bishop, C.M., Bishop, H.: Deep learning: Foundations and concepts. Springer Nature (2023)
- 12. Brauwers, G., Frasincar, F.: A general survey on attention mechanisms in deep learning. Transactions on Knowledge and Data Engineering (2021)
- 13. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems (2020)
- 14. Chen, R.T., Rubanova, Y., Bettencourt, J., Duvenaud, D.K.: Neural ordinary differential equations. Advances in neural information processing systems 31 (2018)
- 15. Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for contrastive learning of visual representations. In: Int. Conf. on machine learning (2020)
- 16. Child, R., Gray, S., Radford, A., Sutskever, I.: Generating long sequences with sparse transformers. arXiv:1904.10509 (2019)
- 17. Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S., et al.: Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311 (2022)
- 18. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv:1810.04805 (2018)
- 19. Dong, S., Wang, P., Abbas, K.: A survey on deep learning and its applications. Computer Science Review (2021)
- 20. Dong, X., Shen, J.: Triplet loss in siamese network for object tracking. In: European Conf. on computer vision (ECCV) (2018)
- 21. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929 (2020)
- 22. Dubey, S.R., Singh, S.K., Chaudhuri, B.B.: Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing 503 , 92-108 (2022)
- 23. Dwivedi, V.P., Bresson, X.: A generalization of transformer networks to graphs. arXiv:2012.09699 (2020)
- 24. Ericsson, L., Gouk, H., Loy, C.C., Hospedales, T.M.: Self-supervised representation learning: Introduction, advances, and challenges. Signal Processing Magazine (2022)
- 25. Foret, P., Kleiner, A., Mobahi, H., Neyshabur, B.: Sharpness-aware minimization for efficiently improving generalization. arXiv:2010.01412 (2020)
- 26. Fu, D., Arora, S., Grogan, J., Johnson, I., Eyuboglu, E.S., Thomas, A., Spector, B., Poli, M., Rudra, A., Ré, C.: Monarch mixer: A simple sub-quadratic gemm-based architecture. Advances in Neural Information Processing Systems 36 (2024)
- 27. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Ré, C.: Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052 (2022)
- 28. Ghiasi, G., Lin, T.Y., Le, Q.V.: Dropblock: A regularization method for convolutional networks. Advances in neural information processing systems (2018)
- 29. Goodfellow, I., Bengio, Y., Courville, A.: Deep learning (2016)
- 30. Google: Palm 2 technical report. https://ai.google/static/documents/ palm2techreport.pdf (2023)
- 31. Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al.: Bootstrap your own latent-a new approach to self-supervised learning. Adv. in neural information processing systems (2020)
- 32. Grover, A., Leskovec, J.: node2vec: Scalable feature learning for networks. In: ACM SIGKDD Int. Conf. on Knowledge discovery and data mining (2016)
- 33. Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 (2023)
- 34. Gu, A., Goel, K., Ré, C.: Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396 (2021)
- 35. Gu, A., Johnson, I., Goel, K., Saab, K., Dao, T., Rudra, A., Ré, C.: Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems 34 , 572-585 (2021)
- 36. Guo, M.H., Lu, C.Z., Liu, Z.N., Cheng, M.M., Hu, S.M.: Visual attention network. arXiv:2202.09741 (2022)
- 37. Guo, M.H., Xu, T.X., Liu, J.J., Liu, Z.N., Jiang, P.T., Mu, T.J., Zhang, S.H., Martin, R.R., Cheng, M.M., Hu, S.M.: Attention mechanisms in computer vision: A survey. Computational Visual Media (2022)
- 38. Hamilton, K., Nayak, A., Boži´ c, B., Longo, L.: Is neuro-symbolic ai meeting its promises in natural language processing? a structured review. Semantic Web (Preprint), 1-42 (2022)
- 39. Han, K., Wang, Y., Chen, H., Chen, X., Guo, J., Liu, Z., Tang, Y., Xiao, A., Xu, C., Xu, Y., et al.: A survey on vision transformer. transactions on pattern analysis and machine intelligence (2022)
- 40. Hasani, R., Lechner, M., Amini, A., Rus, D., Grosu, R.: Liquid time-constant networks. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 35, pp. 7657-7666 (2021)
- 41. He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Conf. on computer vision and pattern recognition (2020)
- 42. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Conf. on computer vision and pattern recognition (2016)
- 43. Hendrycks, D., Gimpel, K.: Gaussian error linear units (gelus). arXiv:1606.08415 (2016)
- 44. Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems (2017)
- 45. Hinton, G.E., Krizhevsky, A., Wang, S.D.: Transforming auto-encoders. In: Artificial Neural Networks and Machine Learning-ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Proceedings, Part I 21. pp. 44-51 (2011)
- 46. Hinton, G.E., Sabour, S., Frosst, N.: Matrix capsules with em routing. In: International conference on learning representations (2018)
- 47. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Conf. on computer vision and pattern recognition (2017)
- 48. Ioffe, S., Szegedy, C.: Batch normalization: Accelerating deep network training by reducing internal covariate shift. In: Int. Conf. on machine learning (2015)
- 49. Izmailov, P., Podoprikhin, D., Garipov, T., Vetrov, D., Wilson, A.G.: Averaging weights leads to wider optima and better generalization. arXiv:1803.05407 (2018)
- 50. Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., et al.: Mixtral of experts. arXiv preprint arXiv:2401.04088 (2024)
- 51. Kalman, R.E.: A new approach to linear filtering and prediction problems (1960)
- 52. Kambhampati, S., Valmeekam, K., Guan, L., Stechly, K., Verma, M., Bhambri, S., Saldyt, L., Murthy, A.: Llms can't plan, but can help planning in llm-modulo frameworks. arXiv preprint arXiv:2402.01817 (2024)
- 53. Kar, K., Kubilius, J., Schmidt, K., Issa, E.B., DiCarlo, J.J.: Evidence that recurrent circuits are critical to the ventral stream's execution of core object recognition behavior. Nature neuroscience 22 (6), 974-983 (2019)
- 54. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Conf. on computer vision and pattern recognition (2020)
- 55. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Conf. on computer vision and pattern recognition (2020)
- 56. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artificial intelligence review (2020)
- 57. Khan, S., Naseer, M., Hayat, M., Zamir, S.W., Khan, F.S., Shah, M.: Transformers in vision: A survey. ACM computing surveys (CSUR) (2022)
- 58. Khoshraftar, S., An, A.: A survey on graph representation learning methods. ACM Transactions on Intelligent Systems and Technology 15 (1), 1-55 (2024)
- 59. Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P., Maschinot, A., Liu, C., Krishnan, D.: Supervised contrastive learning. Advances in neural information processing systems (2020)
- 60. Kiefer, J., Wolfowitz, J.: Stochastic estimation of the maximum of a regression function. The Annals of Mathematical Statistics (1952)
- 61. Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv:1412.6980 (2014)
- 62. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional networks. arXiv:1609.02907 (2016)
- 63. Lewis, M., Liu, Y., Goyal, N., Ghazvininejad, M., Mohamed, A., Levy, O., Stoyanov, V., Zettlemoyer, L.: BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 7871-7880 (2020)
- 64. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V ., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al.: Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems 33 , 9459-9474 (2020)
- 65. Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object detection. In: Int. Conf. on computer vision (2017)
- 66. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., Han, J.: On the variance of the adaptive learning rate and beyond. arXiv:1908.03265 (2019)
- 67. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., Neubig, G.: Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys (2023)
- 68. Liu, P.J., Saleh, M., Pot, E., Goodrich, B., Sepassi, R., Kaiser, L., Shazeer, N.: Generating wikipedia by summarizing long sequences. arXiv:1801.10198 (2018)
- 69. Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv:1907.11692 (2019)
- 70. Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Int. Conf. on computer vision (2021)
- 71. Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. arXiv:1711.05101 (2017)
- 72. Maass, W.: Networks of spiking neurons: the third generation of neural network models. Neural networks 10 (9), 1659-1671 (1997)
- 73. Mandar Kahade: Gpt-4: 8 models in one. https://www.kdnuggets.com/2023/08/ gpt4-8-models-one-secret.html (2023), accessed: 2024-07-20
- 74. Mao, J., Gan, C., Kohli, P., Tenenbaum, J.B., Wu, J.: The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584 (2019)
- 75. Mescheder, L., Geiger, A., Nowozin, S.: Which training methods for GANs do actually converge? In: Int. Conf. on machine learning (2018)
- 76. Meta: The llama 3 herd of models. Online (2024), https://ai.meta.com/research/ publications/the-llama-3-herd-of-models/ , accessed: 2024-07-19
- 77. Min, B., Ross, H., Sulem, E., Veyseh, A.P.B., Nguyen, T.H., Sainz, O., Agirre, E., Heinz, I., Roth, D.: Recent advances in natural language processing via large pre-trained language models: A survey. arXiv preprint arXiv:2111.01243 (2021)
- 78. Misra, D.: Mish: A self regularized non-monotonic activation function. arXiv:1908.08681 (2019)
- 79. Mnih, V., Badia, A.P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., Silver, D., Kavukcuoglu, K.: Asynchronous methods for deep reinforcement learning. In: Int. Conf. on machine learning (2016)
- 80. Moradi, R., Berangi, R., Minaei, B.: A survey of regularization strategies for deep models. Artificial Intelligence Review (2020)
- 81. Nair, V., Hinton, G.E.: Rectified linear units improve restricted boltzmann machines. In: Int. Conf. on machine learning (ICML-) (2010)
- 82. OpenAI: Chatgpt: Optimizing language models for dialogue. https://openai.com/ blog/chatgpt/ (2022)
- 83. OpenAI: Gpt-4 technical report (2023)
- 84. OpenAI: Hello gpt-4o! https://openai.com/index/hello-gpt-4o/ (2024), accessed: 2024-07-19
- 85. Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al.: Training language models to follow instructions with human feedback. arXiv:2203.02155 (2022)
- 86. Parisotto, E., Mohamed, A.r., Singh, R., Li, L., Zhou, D., Kohli, P.: Neuro-symbolic program synthesis. arXiv preprint arXiv:1611.01855 (2016)
- 87. Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., et al.: Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048 (2023)
- 88. Poli, M., Massaroli, S., Nguyen, E., Fu, D.Y., Dao, T., Baccus, S., Bengio, Y., Ermon, S., Ré, C.: Hyena hierarchy: Towards larger convolutional language models. In: International Conference on Machine Learning. pp. 28043-28078 (2023)
- 89. Radford, A., Narasimhan, K., Salimans, T., Sutskever, I., et al.: Improving language understanding by generative pre-training (2018)
- 90. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog (2019)
- 91. Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36 (2024)
- 92. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research (2020)
- 93. Ramachandran, P., Zoph, B., Le, Q.V.: Searching for activation functions. arXiv preprint arXiv:1710.05941 (2017)
- 94. Reddi, S.J., Kale, S., Kumar, S.: On the convergence of adam and beyond. arXiv:1904.09237 (2019)
- 95. Reid, M., Savinov, N., Teplyashin, D., Lepikhin, D., Lillicrap, T., Alayrac, J.b., Soricut, R., Lazaridou, A., Firat, O., Schrittwieser, J., et al.: Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024)
- 96. Sabour, S., Frosst, N., Hinton, G.E.: Dynamic routing between capsules. Advances in neural information processing systems 30 (2017)
- 97. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C.: Mobilenetv2: Inverted residuals and linear bottlenecks. In: Conf. on computer vision and pattern recognition (2018)
- 98. Sanh, V., Debut, L., Chaumond, J., Wolf, T.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv:1910.01108 (2019)
- 99. de Santana Correia, A., Colombini, E.L.: Attention, please! a survey of neural attention models in deep learning. Artificial Intelligence Review 55 (8), 6037-6124 (2022)
- 100. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., Scialom, T.: Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems 36 (2024)
- 101. Schneider, J., Meske, C., Kuss, P.: Foundation models: a new paradigm for artificial intelligence. Business & Information Systems Engineering pp. 1-11 (2024)
- 102. Schneider, J., Vlachos, M.: Reflective-net: Learning from explanations. Data Mining and Knowledge Discovery pp. 1-22 (2023)
- 103. Schneider, J., Vlachos, M.: A survey of deep learning: From activations to transformers. In: Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART) (2024)
- 104. Schneider, J., Wattenhofer, R.: Trading bit, message, and time complexity of distributed algorithms. In: International Symposium on Distributed Computing. pp. 51-65. Springer (2011)
- 105. Schultz, M., Joachims, T.: Learning a distance metric from relative comparisons. Advances in neural information processing systems 16 (2003)
- 106. Shao, J., Hu, K., Wang, C., Xue, X., Raj, B.: Is normalization indispensable for training deep neural network? Advances in Neural Information Processing Systems (2020)
- 107. Shazeer, N.: Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150 (2019)
- 108. Shazeer, N., Stern, M.: Adafactor: Adaptive learning rates with sublinear memory cost. In: Int. Conf. on Machine Learning (2018)
- 109. Shrestha, A., Mahmood, A.: Review of deep learning algorithms and architectures. IEEE access (2019)
- 110. Sieber, J., Alonso, C.A., Didier, A., Zeilinger, M.N., Orvieto, A.: Understanding the differences in foundation models: Attention, state space models, and recurrent neural networks. arXiv preprint arXiv:2405.15731 (2024)
- 111. Skinner, B.F.: Reinforcement today. American Psychologist 13 (3), 94 (1958)
- 112. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L.: Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems (2020)
- 113. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research (2014)
- 114. Sun, R.Y.: Optimization for deep learning: An overview. Operations Research Society of China (2020)
- 115. Sun, Y., Dong, L., Huang, S., Ma, S., Xia, Y., Xue, J., Wang, J., Wei, F.: Retentive network: A successor to transformer for large language models. arXiv preprint arXiv:2307.08621 (2023)
| 116. | Sun, Z., Deng, Z.H., Nie, J.Y., Tang, J.: Rotate: Knowledge graph embedding by relational rotation in complex space. arXiv:1902.10197 (2019) |
|--------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| 117 | Team, G., Anil, R., Borgeaud, S., Wu, Y., Alayrac, J.B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A.M., Hauth, A., et al.: Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 (2023) |
| 118 | Tolstikhin, I.O., Houlsby, N., Kolesnikov, A., Beyer, L., Zhai, X., Unterthiner, T., Yung, J., Steiner, A., Keysers, D., Uszkoreit, J., et al.: Mlp-mixer: An all-mlp architecture for vision. Advances in neural information processing systems 34 , 24261-24272 (2021) |
| 119 | Touvron, H., Cord, M., Sablayrolles, A., Synnaeve, G., Jégou, H.: Going deeper with image transformers. In: Int. Conf. on Computer Vision (2021) |
| 120 | Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) |
| 121 | Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al.: Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023) |
| 122 | Ulyanov, D., Vedaldi, A., Lempitsky, V.: Instance normalization: The missing ingredient for fast stylization. arXiv:1607.08022 (2016) |
| 123 | Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information processing sys- tems (2017) |
| 124 | Veliˇkovi´, c c P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv:1710.10903 (2017) |
| 125 | Wang, F., Jiang, M., Qian, C., Yang, S., Li, C., Zhang, H., Wang, X., Tang, X.: Residual attention network for image classification. In: Conf. on computer vision and pattern recog- nition (2017) |
| 126 | Wang, Q., Ma, Y., Zhao, K., Tian, Y.: Acomprehensive survey of loss functions in machine learning. Annals of Data Science (2020) |
| 127 | Williams, R.J., Peng, J.: Function optimization using connectionist reinforcement learning algorithms. Connection Science (1991) |
| 128 | Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., Philip, S.Y.: A comprehensive survey on graph neural networks. Transactions on neural networks and learning systems (2020) |
| 129 | Xie, Q., Luong, M.T., Hovy, E., Le, Q.V.: Self-training with noisy student improves ima- genet classification. In: Conf. on computer vision and pattern recognition (2020) |
| 130 | Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks. In: Conf. on computer vision and pattern recognition (2017) |
| 131 | Yang, X., Song, Z., King, I., Xu, Z.: A survey on deep semi-supervised learning. Transac- tions on Knowledge and Data Engineering (2022) |
| 132 | Yang, Z., Dai, Z., Yang, Y., Carbonell, J., Salakhutdinov, R.R., Le, Q.V.: Xlnet: General- ized autoregressive pretraining for language understanding. Advances in neural information processing systems (2019) |
| 133 | You, Y., Li, J., Reddi, S., Hseu, J., Kumar, S., Bhojanapalli, S., Song, X., Demmel, J., Keutzer, K., Hsieh, C.J.: Large batch optimization for deep learning: Training bert in 76 minutes. arXiv:1904.00962 (2019) |
| 134 | Yuksel, S.E., Wilson, J.N., Gader, P.D.: Twenty years of mixture of experts. IEEE transac- |
| 135 | tions on neural networks and learning systems 23 (8), 1177-1193 (2012) Zbontar, J., Jing, L., Misra, I., LeCun, Y., Deny, S.: Barlow twins: Self-supervised learning via redundancy reduction. In: Int. Conf. on Machine Learning (2021) |
| 136 | Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle- consistent adversarial networks. In: Int. Conf. on computer vision (2017) | | null | [
"Johannes Schneider"
] | 2024-08-01T08:50:25+00:00 | 2024-08-01T08:50:25+00:00 | [
"cs.LG"
] | What comes after transformers? -- A selective survey connecting ideas in deep learning | Transformers have become the de-facto standard model in artificial
intelligence since 2017 despite numerous shortcomings ranging from energy
inefficiency to hallucinations. Research has made a lot of progress in
improving elements of transformers, and, more generally, deep learning
manifesting in many proposals for architectures, layers, optimization
objectives, and optimization techniques. For researchers it is difficult to
keep track of such developments on a broader level. We provide a comprehensive
overview of the many important, recent works in these areas to those who
already have a basic understanding of deep learning. Our focus differs from
other works, as we target specifically novel, alternative potentially
disruptive approaches to transformers as well as successful ideas of recent
deep learning. We hope that such a holistic and unified treatment of
influential, recent works and novel ideas helps researchers to form new
connections between diverse areas of deep learning. We identify and discuss
multiple patterns that summarize the key strategies for successful innovations
over the last decade as well as works that can be seen as rising stars.
Especially, we discuss attempts on how to improve on transformers covering
(partially) proven methods such as state space models but also including
far-out ideas in deep learning that seem promising despite not achieving
state-of-the-art results. We also cover a discussion on recent state-of-the-art
models such as OpenAI's GPT series and Meta's LLama models and, Google's Gemini
model family. |
2408.00387v1 | ## Decomposition of Nonlinear Collision Operator in Quantum Lattice Boltzmann Algorithm
E. Dinesh Kumar ∗ and Steven H. Frankel †
Faculty of Mechanical Engineering
Technion - Israel Institute of Technology, Haifa, 3200003, Israel. ∗ †
[email protected] , [email protected]
Abstract - We propose a quantum algorithm to tackle the quadratic nonlinearity in the Lattice Boltzmann (LB) collision operator. The key idea is to build the quantum gates based on the particle distribution functions (PDF) within the coherence time for qubits. Thus, both the operator and a state vector are linear functions of PDFs, and upon quantum state evolution, the resulting PDFs will have quadraticity. To this end, we decompose the collision operator for a DmQn lattice model into a product of 2( n +1) operators, where n is the number of lattice velocity directions. After decomposition, the ( n +1) operators with constant entries remain unchanged throughout the simulation, whereas the remaining ( n +1) will be built based on the statevector of the previous time step. Also, we show that such a decomposition is not unique. Compared to the second-order Carleman-linearized LB, the present approach reduces the circuit width by half and circuit depth by exponential order. The proposed algorithm has been verified through the one-dimensional flow discontinuity and two-dimensional Kolmogrov-like flow test cases.
Introduction. -The potential for simulating fluid flows on quantum computers has generated significant interest due to the unique properties of qubits, such as superposition and entanglement. By encoding flow variables onto the probability amplitudes of qubit states, we can achieve significantly reduced memory requirements and a potentially exponential increase in computing speed. However, quantum operations are inherently linear and unitary, while the equations of fluid mechanics exhibit nonlinearity. As a result, existing computational fluid dynamics (CFD) algorithms need to be redesigned and adapted to fully utilize the power of quantum processor units (QPUs).
Traditional CFD methods solve the discretized NavierStokes equation (NSE), and commonly used algorithms are implicit in time-stepping, which requires linear system solvers (LSS) [1-5]. The Lattice Boltzmann Method (LBM) has been considered an alternate approach to NSEbased CFD solvers and has been studied extensively for various flow problems. Due to locality in space and explicit time-stepping, LBM does not require LSS and has proven to be efficient on graphic processor units (GPUs) [6]. Compared to NSE, where the nonlinear term ( u · ∇ u ) is also non-local, the LB equation has nonlinearity ( u · u ) in the equilibrium function and is local. This nonlinearity does not pose any difficulty in classical computers, as u will be saved as a floating point number followed by a square operation [7,8]. However, if we mimic the classical procedure on QPUs, the quantum advantage may be lost, especially when amplitude encoding is chosen.
There are several existing works on Quantum Lattice Boltzmann (QLB) for fluid flow simulation, but most of them ignore nonlinearity and are only used for very low Reynolds number cases [9-13]. In the absence of nonlinearity, the particle distribution functions (PDF) f can be encoded directly as probability amplitudes of qubits. However, evaluating the nonlinear term involves the vector f ⊗ f . [14] highlighted that it is impossible to have a linear operator that performs the squares of probability amplitudes, as the resulting state vector may not be a unit vector. Furthermore, the no-cloning theorem states that it is impossible to copy an arbitrary quantum state to create an identical quantum state [14]. To address this, the Carleman linearization (CL) technique can be utilized which converts the finite-dimensional nonlinear system into a infinite-dimensional linear system which is then truncated for implementation [15]. Previously, CL has been successfully applied to LBM up to second-order truncation for low to moderate Reynolds number cases. When CL is applied to LBM, the resultant state vector becomes infinite, i.e., ( f f , ⊗ f f , ⊗ f ⊗ f , . . . ). In addition, another approximation has to be made for the term 1 ρ ≈ 2 -ρ , where ρ is the density, which is valid for weakly incompressible regimes. Even after successive simplifications, the Carleman LB (CLB) algorithm still requires repeated encoding and read-out of state vectors for each time step [16-19].
In the present work, instead of the CL algorithm, we propose the decomposition of the LB operator as a product of multiple operators derived from the PDFs from the previous time step. We retain the first-order Taylor ap-
proximation for 1 ρ and accept that the encoding and readout process is required for each time step [17-19]. Our approach does not involve any truncation of infinite sets. Instead, the state vector is a finite set, specifically, four sets of f appended sequentially.
In the following, we briefly describe the LB equations, followed by deriving the quadratic formulation of the LB collision operation. Later, we will present the quantum algorithm and the computational complexity of two-qubit gates.
Lattice Boltzmann Method. LBM describes the evolution of flow field through PDFs in a uniform Cartesian grid with the chosen lattice model DmQn , where m and n denotes the number of spatial dimensions and velocity directions, respectively. The single relaxation time LB model for the flow field is given by
$$f _ { i } ( \mathbf x + \mathbf e _ { i } \Delta t, t + \Delta t ) = f _ { i } ( \mathbf x, t ) - \frac { \Delta t } { \tau } \left [ f _ { i } ( \mathbf x, t ) - f _ { i } ^ { e q } ( \mathbf x, t ) \right ] \quad \text{and} \ \delta _ { j k } \\ \text{expressi} \\ \text{$\Phi$} \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Psi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi\Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phis \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi \Phi$$
where f i represents the PDF along the i th direction, e i is the lattice velocity, and the kinematic viscosity ν is related to the relaxation parameter τ = 3 ν +0 5. The equilibrium . distribution function (EDF) is given by
$$f _ { i } ^ { e q } = w _ { i } \rho \left [ 1 + \frac { e _ { i } \cdot \mathbf u } { c _ { s } ^ { 2 } } + \frac { ( e _ { i } \cdot \mathbf u ) ^ { 2 } } { 2 c _ { s } ^ { 4 } } - \frac { \mathbf u \cdot \mathbf u } { 2 c _ { s } ^ { 2 } } \right ] \quad ( 2 ) \quad \ f _ { i } ^ { * } \colon$$
where w i is the lattice weight, c s the sound speed, ρ the fluid density, and u is the flow velocity. The zeroth and first moment of PDFs yields the macroscopic flow variables density( ρ ) and momentum ( ρ u ),
$$\rho ( { \mathbf x }, t ) & = \sum _ { i } f _ { i } ( { \mathbf x }, t ) & ( 3 ) \quad \text{when} \\ o ( { \mathbf x }. t ) \mathbf u ( { \mathbf x }. t ) & = \sum \mathbf e _ { i } f _ { i } ( { \mathbf x }. t ) & ( 4 )$$
$$\rho ( x, t ) \mathbf u ( x, t ) = \sum _ { i } ^ { \frac { i } { i } } \mathbf e _ { i } f _ { i } ( x, t ) \quad \quad ( 4 ) \\.. \ \ - \. \ \. \ \. \ \. \ \.. \ \.$$
Typically, LB algorithm is split into collision and streaming steps. In the collision step, the interaction of particles and their relaxation to equilibrium state is modelled. Later the particles move to their neighbouring site according to the lattice velocity, a process termed streaming. Thus, Eq. (1) can be written as
$$f _ { i } ^ { * } ( { \mathbf x }, t ) = f _ { i } ( { \mathbf x }, t ) - \frac { \Delta t } { \tau } \left [ f _ { i } ( { \mathbf x }, t ) - f _ { i } ^ { e q } ( { \mathbf x }, t ) \right ] \quad ( 5 ) \quad \underset { \dots } { \mathbf x } \quad \text{where} }$$
$$f _ { i } ( { \mathbf x }, t + \Delta t ) & = f _ { i } ^ { * } ( { \mathbf x } - { \mathbf e } _ { i } \Delta t, t ) & ( 6 ) \underset { \text{treated} } { \text{ where } }$$
where f ∗ i is the post-collision state. In the presence of a wall boundary, we use half-way bounce back scheme to reconstruct PDFs coming from the wall node ( f -i ),
$$f _ { - i } ( { \mathbf x } _ { b }, t + \Delta t ) = f _ { i } ^ { * } ( { \mathbf x }, t ) \quad \quad ( 7 ) \ \text{ when}$$
Formulation. -
Quadratic form for collision step. First, we rewrite Eq. (2) and using Eqs. (3) and (4), we get the quadratic form for EDF as,
$$\text{sets}. \quad & \text{form for EDF as}, \\ \text{four} \\ \text{ions}, \quad & \text{f} _ { i } ^ { e q } = \frac { 1 } { \rho } w _ { i } \left [ \rho ^ { 2 } + \rho \frac { e _ { i } \cdot \rho \mathbf u } { c _ { s } ^ { 2 } } + \frac { ( e _ { i } \cdot \rho \mathbf u ) ^ { 2 } } { 2 c _ { s } ^ { 4 } } - \frac { \rho \mathbf u \cdot \rho \mathbf u } { 2 c _ { s } ^ { 2 } } \right ] \quad ( 8 ) \\ \text{e LB} \\ \text{return} \\ \text{public} \quad & \text{$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$\quad$
\quad \cdot \quad \cdot \quad \sim \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot \quad \cdot$$
where the coefficient α ijk is given by
$$\mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } \\ \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } \\ \mathfrak { d } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s }& \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s } & \mathfrak { s }$$
and δ jk denotes the Kronecker delta function. Next, the expression for collision in Eq. (5) can be rewritten as,
$$\dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \dots \delta$$
Using Eqs. (3) and (10), the quadratic form of the collision step can be written as
$$& \text{step can be wrtten as} \\ \text{the} \\ \text{and} \\ \text{ables}$$
where β ijk = γ ijk · ( 1 -∆ t τ ) + χ jk · ∆ t τ α ijk and T denotes the transpose. The indicator functions γ and χ for i th lattice direction is given by,
$$\gamma _ { _ { i j k } } = \delta _ { i j } + \delta _ { i k } - \delta _ { i j } \cdot \delta _ { i k } \quad \quad ( 1 3 )$$
$$\gamma _ { i j k } & = \delta _ { i j } + \delta _ { i k } - \delta _ { i j } \cdot \delta _ { i k } \quad \quad ( 1 3 ) \\ \chi _ { j k } & = \begin{cases} 1 & \text{if $j\leq k$} \\ 0 & \text{otherwise} \end{cases}$$
Matrix formulation for weakly compressible LBM. In case of weakly compressible flows, ρ -1 ≈ 2 -ρ , we then express
$$\int _ { \ell = 1 } ( 2 - \rho ) = ( - 1, - 1, \cdots, - 1, 2 ) \cdot ( f _ { 1 }, f _ { 2 }, \cdots, f _ { n _ { \ell } }, 1 ) ^ { \mathrm T } \ ( 1 5 )$$
where the value 1 appended at the end of PDFs can be treated as a auxiliary constant. Therefore, the collision step can be written as
$$\intertext { e t o } \, \begin{pmatrix} \quad \,. \\ \quad \, \left ( \begin{smallmatrix} f _ { i } ^ { * } \\ 1 \end{smallmatrix} \right ) = \left ( \begin{smallmatrix} f \\ 1 \end{smallmatrix} \right ) ^ { \mathrm T } \left ( \begin{smallmatrix} \beta _ { i } & 0 \\ 0 & 1 \end{smallmatrix} \right ) \left ( \begin{smallmatrix} \text{Diag} ( f ) & 0 \\ 0 & 1 \end{smallmatrix} \right ) \, \mathrm W \left ( \begin{smallmatrix} f \\ 1 \end{smallmatrix} \right ) \, \end{pmatrix} \, \begin{smallmatrix} \quad \, ( 1 6 ) \\ \quad \, \left ( \begin{smallmatrix} \quad \,. \\ 1 \end{smallmatrix} \right ) \, \end{pmatrix}$$
where f = ( f , f 1 2 , · · · , f n e ), and β i represent the upper triangular matrix in Eq. (12). The term Diag ( f ) represents the diagonal matrix with entries from f . The matrix
W is given by
$$\dot { \ } g \text{ven by } & & & & & & & & & &$$
## Quantum Algorithm. -
Encoding. Let n g denote the total number of grid points and the number of PDFs will be n f = n n e g . According to the lattice directions, we arrange and order the PDFs in a vector form, df = ( f 1 , f 2 , . . . , f n e , 1), where the suffix refers to the direction, and each f i is of length n g . Since PDFs are encoded as a quantum state-vector, we define φ = ( df , df , df , df ). A computational register containing n q = log ( 2 n f ) qubits along with two ancilla qubits will be required to encode the PDFs. The initial state to encode will be
$$\left | \phi ^ { 0 } \right \rangle = \left | 0 \right \$$
Collision. Before we express matrix version of Eq. (16), we define the following matrices:
$$\widetilde { W } & = \begin{pmatrix} - \mathbf I _$$
$$\widetilde { \mathbf F } _ { i } = \begin{pmatrix}$$
$$\begin{array} {$$
where I g rc denotes the identity matrix of size n g × n g , 2 g i = (2 , · · · , 2) and 0 g i = (0 , · · · , 0) are the vectors of size n g . The term ˜ β ric = χ ic β ric I g rc , where β ric is defined in Eq. (12), with r and c represent row and column indices, respectively. Based on the above set of matrices, we
$$\begin{array} { c c c c c c c c } & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & && & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \\ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & - & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & _ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ; & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &. & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & } & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & . & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & ; & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & \ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &. \ & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &.. & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &.
& & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & & &$$
Therefore, the collision operator is given by
| φ ∗ 〉 = ̂ F n e ̂ B n e · · · ̂ F 2 ̂ B F 2 ̂ 1 ̂ B D W 1 ̂ ̂ | φ 0 〉 (25) In Eq. (25), when the SV is multiplied by ̂ W yields the value 2 -ρ , which then multiplied by ̂ D results in the first order approximation of f i ρ , i.e., the column vector defined on the right side of Eq. (12). Upon successive multiplication by ̂ F B i ̂ i , yields the PDFs in the i th direction. Thus, for n e velocity directions, the collision operation is decomposed into 2 n e +2 operations. It is important to note that the state vector has been initialized with four sets of df . Among these, one will ultimately lead to the final outcome, while another is designated for storing temporary values. The remaining two sets remain constant, providing the initial state vector values during intermediate computations. Consequently, the matrices defined in Eq. (24) are not unique. In particular, the following is a valid choice:
$$( 2 1 ) & \quad \text{not unique. In particular, the following is a valid choice:} \\ \widehat { B } _ { i } = \begin{pmatrix} I & 0 & 0 & 0 \\ 0 & I & 0 \\ \widetilde { B } _ { i } & 0 & 0 \\ I \end{pmatrix} ; \ \widehat { \mathbf F } _ { i } = \begin{pmatrix} I & 0 & \widetilde { \mathbf F } _ { i } & 0 \\ 0 & 0 & I \\ I & 0 \\ I \end{pmatrix} \quad ( 2 6 ) \\ \quad \text{Streaming and Boundary Conditions.} \quad \text{The streaming} \\ \quad \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot \ \cdot$$
Streaming and Boundary Conditions. The streaming matrix S consists of n e × n e blocks with the size of each block n g × n g . The matrix S is square binary matrix where each row/column contains one entry equal to 1. Since every row index i can be decomposed as x + yn x + i e n n x y , and the corresponding column index j such that s ij = 1, can be computed by replacing the co-ordinate x by x -e i ∆ . In the present study, we utilized the wall boundary t condition, accordingly the column index will be computed as x ∆ + y ∆ n x + i -e n n x y , where ( x ∆ , y ∆ ) = x -e i and i -e is the reflection of the direction i e . Based on the matrix S , we define the streaming operator as
$$\begin{array} { c c c } S, \, \text{we define the streaming operator as} \\ n _ { g }, \\ \intertext { of } \begin{pmatrix} \widetilde { S } & \\ & I \\ & & I \\ \end{pmatrix} & ; \, \text{where } \widetilde { S } = \begin{pmatrix} S & \\ & 1 \end{pmatrix} \quad ( 2 7 ) \\ \intertext { \. }$$
From Eq. (27), it is clear that streaming is performed only on the computational qubits without any involvement from ancilla qubits. Thus, the PDF for the next time step can be obtained from the quantum state,
## Results and Discussion. -
$$| \phi ^ { n } \rangle = \widehat { S } \, | \phi ^ { * } \rangle \quad \quad ( 2 8 ) \quad \text{$\upsilon_{c}$} \\ \text{Discussion}. \, - \, \quad \ \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{\quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{\quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsi_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$} \quad \text{$\upsilon_{c}$,}$$
Verification - Case 1: 1D discontinuity flow. The classical shock tube discontinuity problem in one dimension is used to verify the present the algorithm. Because of the weak compressibility assumption, we refer it as a discontinuity rather than a shock. The one dimensional domain of size n g = 500 grid points and the D Q 1 3 lattice has been chosen. At the initial time, the velocity is set to zero everywhere and density is defined as,
$$\rho ( x, 0 ) = \begin{cases} 1. 0 + \Delta \rho & \text{if $x\leq \frac{n_{q}}{2}$} \\ 1. 0 & \text{otherwise} \end{cases} \quad ( 2 9 ) \quad a \text{fun} \\ \text{son o}$$
where ∆ ρ = 5 × 10 -5 is chosen to satisfy the weak incompressibility assumption made earlier. Bounce-back boundary condition is applied at the both ends of the domain. After simulating 200 time steps, we obtained a good comparison with the exact Riemann solution for normalized pressure ( p ∗ = ( p -p s ) /p s ; p s = ∆ ρc 2 s ) and velocity ( u ) [20] and is given in Fig. 1. The relative difference between exact and QLB results is to found to be 5%.
Fig. 1: Comparison of normalized pressure and velocity of 1D discontinuity flow obtained from QLB with the exact Riemann solution at 200 th time step.
Case 2: 2D Kolmogorov flow. In order to compare the accuracy of present algorithm with that of Carleman linearized - LBM presented in [19], we performed the twodimensional simulation of Kolmogorov-like flow on a 32 × 32 grid. The D2Q9 lattice is used. Initially, the PDFs are defined as,
$$f _ { i } ( x, y ) = w _ { i } \left [ 1 + A _ { x } \cos \left ( \frac { 2 \pi k _ { x } } { N _ { y } } y \right ) \text{e} _ { i } \cdot \text{e} _ { x } + \text{the $rc$} \text{The r} \\ A _ { y } \cos \left ( \frac { 2 \pi k _ { y } } { N _ { x } } x \right ) \text{e} _ { i } \cdot \text{e} _ { y } \right ] \text{ } ( 3 0 ) \text{ the stat} \\ LB \text{ ope}$$
where e x = (1 0) and , e y = (0 1). , The parameters in Eq. (30) are taken as: A x = 0 3, . A y = 0 2, . k x = 1, and k y = 4. We choose different viscosity values ν ranging from 1 / 6 to 0 0088, and for each . ν we ran the simulation up to 100 time steps. After simulating 100 time steps, we compute the mean value of Root Mean Squared Error (RMSE) between the distribution functions obtained from the present algorithm ( f p i ) and the exact LBM simulation ( f e i ),
$$\stackrel { \nabla _ { i } \, \jmath } { s } - \bigvee _ { i } \nabla _ { i } \, \jmath, \\ \text{is} \, \ e \, \begin{array} { c } \nabla _ { i } \\ \nabla _ { i } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end {array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \, \left \langle \begin{array} { c } \nabla _ { j } \end{array} \right \rangle$$
Since CL transforms a finite set to an infinite dimensional set of equations, [19] analysed two different approaches, truncation (neglecting higher order terms) and closure (approximate the product of functions into products of a function and a constant). Fig. 2 shows the comparison of RSME with the results of Carleman second order truncation and closure approaches presented in [19]. Since the present approach does not require any order of truncation except the assumption on weak compressiblity, the < RSME > value is below 10 -5 compared to 10 -2 from Carleman closure approach.
Fig. 2: Comparison of root mean square error obtained for various viscosity values are compared with thre results of carleman second order truncation and closure approaches presented in [19].
Computational Complexity. In the present approach, the number of qubits required will be 2+log 2 ( n f ), whereas the Carleman truncated system of k th order would require 1+log ( 2 n f + n 2 f + . . . + n k f ) qubits. Fig. 3a shows the number of qubits required for the grid size ranging from 10 1 to 10 20 using D2Q9 lattice. For larger grid sizes, the CLsecond (CL2) and third (CL3) order truncation approach requires almost twice and thrice the qubit resource than the present approach.
The number of two-qubit gates is one of the ways to estimate the complexity of the present quantum algorithm. Since PDFs are encoded as amplitudes of the qubit states, the state preparation step is required before performing LB operations. According to [21], O (2 n q ) CNOT gates
are required for state preparation. Based on Eq. (25) and (28), the number of LB operators per time step is 2 n e +3, and each operator would require O (2 n q -1 (2 n q -1)) CNOT gates. In the second-order Carleman truncation, the qubit requirement will be n c q = 1 + log ( 2 n f + n 2 f ). Correspondingly, the O (2 n c q -1 (2 n c q -1)) CNOT gates will be required. Fig. 3b shows the number of CNOT gates required for different grid sizes. For smaller grids of size 10 3 , the gate counts of CL2 are a few orders higher than the present approach; however, for the larger grid, it is exponentially larger.
nCNOT
Fig. 3: Comparison of (a) number of qubits and (b) number of CNOT gates required for number of grid points between the present algorithm vs. the Carleman second and third order truncation.
Summary. We have developed a quantum algorithm to deal with the non-linearity in the Lattice Boltzmann collision operator. By decomposing the collision operator, we significantly reduced the circuit width by half and circuit depth in exponential order when compared to the Carleman linearization technique. While there are still challenges to be addressed, such as the encoding and readout process for each time step and the construction of quantum gates, this work aims to minimize the quantum resources required by the CL technique and prevent the formation of an infinite system. Due to the weakly compressible assumption and the first-order approximation for 1 /ρ , this work is limited to moderate Reynolds numbers of O(100).
Acknowledgments. This research was kindly supported by the Quantum Computing Consortium that has been funded by the MAGNET program of the Israel Innovation Authority (IIA).
## REFERENCES
- [1] Harrow A. W., Hassidim A. and Lloyd S. , Physical Review Letters , 103 (2009) .
- [2] Cao Y., Anmer D., Frankel S. H. and Sabre K. , Molecular Physics , 110 (2012) 1675.
- [3] Bharadwaj S. S. and Sreenivasan K. R. , Proceedings of the National Academy of Sciences , 120 (2023) e2311014120.
- [4] Ingelmann J., Bharadwaj S. S., Pfeffer P., Sreenivasan K. R. and Schumacher J. , Computers & Fluids , (2024) 106369.
- [5] Bharadwaj S. S. and Sreenivasan K. R. , arXiv preprint arXiv:2405.09767 , (2024) .
- [6] Kuznik F., Obrecht C., Rusaouen G. and Roux J.-J. , Computers & Mathematics with Applications , 59 (2010) 2380.
- [7] Steijl R. , Applied Sciences 2023, Vol. 13, Page 529 , 13 (2022) 529.
- https://www.mdpi.com/2076-3417/13/1/529/htmhttps://www.mdpi
- [8] Steijl R. , , (2024) .
- [9] Mezzacapo A., Sanz M., Lamata L., Egusquiza I. L., Succi S. and Solano E. , Scientific Reports 2015 5:1 , 5 (2015) 1.
## https://www.nature.com/articles/srep13153
- [10] Budinski L. , Quantum Information Processing , 20 (2021) 57.
- [11] Budinski L. , Int. J. Quantum Information , 20 (2022) 2150039.
- [12] Schalkers M. A. and M¨ oller M. , J. Comp. Phys. , 502 (2024) 112816.
- [13] Dinesh Kumar E. and Frankel S. H. , arXiv:2405.08669 , (2024) .
- [14] Succi S., Itani W., Sreenivasan K. and Steijl R. , Europhysics Letters , 144 (2023) 10001.
https://dx.doi.org/10.1209/0295-5075/acfdc7
- [15] Carleman T. , , (1932) .
- [16] Itani W. and Succi S. , Fluids , 7 (2022) 24.
- [17] Itani W., Sreenivasan K. R. and Succi S. , Phys. Fluids , 36 (2024) 017112.
## https://doi.org/10.1063/5.0176569
- [18] Sanavio C., Scatamacchia R., de Falco C. and Succi S. , Physics of Fluids , 36 (2024) .
- [19] Sanavio C. and Succi S. , AVS Quantum Science , 6 (2024) 023802.
https://doi.org/10.1116/5.0195549
- [20] Toro E. F. , Riemann solvers and numerical methods for fluid dynamics: a practical introduction (Springer Science & Business Media) 2013.
- [21] Shende V. V., Bullock S. S. and Markov I. L. , Synthesis of quantum logic circuits in proc. of Proceedings of the 2005 Asia and South Pacific Design Automation Conference 2005 pp. 272-275. | 10.1209/0295-5075/ad8bef | [
"Dinesh Kumar E",
"Steven H. Frankel"
] | 2024-08-01T08:56:38+00:00 | 2024-08-01T08:56:38+00:00 | [
"quant-ph"
] | Decomposition of Nonlinear Collision Operator in Quantum Lattice Boltzmann Algorithm | We propose a quantum algorithm to tackle the quadratic nonlinearity in the
Lattice Boltzmann (LB) collision operator. The key idea is to build the quantum
gates based on the particle distribution functions (PDF) within the coherence
time for qubits. Thus, both the operator and a state vector are linear
functions of PDFs, and upon quantum state evolution, the resulting PDFs will
have quadraticity. To this end, we decompose the collision operator for a
$DmQn$ lattice model into a product of $2(n+1)$ operators, where $n$ is the
number of lattice velocity directions. After decomposition, the $(n+1)$
operators with constant entries remain unchanged throughout the simulation,
whereas the remaining $(n+1)$ will be built based on the statevector of the
previous time step. Also, we show that such a decomposition is not unique.
Compared to the second-order Carleman-linearized LB, the present approach
reduces the circuit width by half and circuit depth by exponential order. The
proposed algorithm has been verified through the one-dimensional flow
discontinuity and two-dimensional Kolmogrov-like flow test cases. |
2408.00388v2 | ## Deepfake Media Forensics: State of the Art and Challenges Ahead
Irene Amerini , Mauro Barni , Sebastiano Battiato , Paolo Bestagini , Giulia 5 8 1 6 Boato , Tania Sari Bonaventura , Vittoria Bruni , Roberto Caldelli , 9 5 5 2 Francesco De Natale , Rocco De Nicola 9 10 , Luca Guarnera , Sara Mandelli , 1 6 Gian Luca Marcialis , Marco Micheletto , Andrea Montibeller , Giulia Orrù , 4 4 9 4 Alessandro Ortis , Pericle Perazzo , Giovanni Puglisi 1 3 4 , Davide Salvi , Stefano 6 Tubaro , Claudia Melis Tonti , Massimo Villari , Domenico Vitulano 6 5 7 5
1 University of Catania, 2 CNIT, Florence, and Universitas Mercatorum, 3 University of Pisa, 4 University of Cagliari, 5 Sapienza University of Rome, 6 Politecnico di Milano, 7 University of Messina, 8 Università di Siena, 9 University of Trento, 10 Scuola IMT Alti Studi Lucca
Abstract. AI-generated synthetic media, also called Deepfakes, have significantly influenced so many domains, from entertainment to cybersecurity. Generative Adversarial Networks (GANs) and Diffusion Models (DMs) are the main frameworks used to create Deepfakes, producing highly realistic yet fabricated content. While these technologies open up new creative possibilities, they also bring substantial ethical and security risks due to their potential misuse. The rise of such advanced media has led to the development of a cognitive bias known as Impostor Bias, where individuals doubt the authenticity of multimedia due to the awareness of AI's capabilities. As a result, Deepfake detection has become a vital area of research, focusing on identifying subtle inconsistencies and artifacts with machine learning techniques, especially Convolutional Neural Networks (CNNs). Research in forensic Deepfake technology encompasses five main areas: detection, attribution and recognition, passive authentication, detection in realistic scenarios, and active authentication. This paper reviews the primary algorithms that address these challenges, examining their advantages, limitations, and future prospects.
Keywords: Multimedia Forensics · Deepfakes.
## 1 Introduction
The advent of Deepfakes, synthetic media generated by Artificial Intelligence (AI) that mimics real images, audio, and video, has significantly impacted various domains including entertainment, politics, and cybersecurity. Deepfakes leverage deep learning techniques, particularly GANs [30] and DMs [39], to create highly convincing but falsified representations of individuals. While these technologies offer creative opportunities, they also pose serious ethical and security challenges due to their potential for misuse. The emergence of such advanced AI-generated
media has led to the development of a cognitive bias known as the Impostor Bias [15], which refers to the tendency to doubt the veracity of multimedia elements due to the knowledge that they can be realistically generated by AI models. Deepfake detection has become an essential field of research, aiming to develop methods to distinguish between real and artificially generated media. Techniques for Deepfake detection often involve analyzing inconsistencies and artifacts that are not easily perceptible to the human eye [36,22] but can be detected using proper detectors based on machine learning algorithms. These detection methods typically focus on both spatial and temporal anomalies in the data, utilizing Convolutional Neural Networks (CNNs) [84,34] for enhanced accuracy. Starting with the Deepfake detection task, the scientific community has over the years taken on several other new challenges to study the nature of synthetic data in detail. We can therefore distinguish 5 main areas of research in the Forensic Deepfake domain, namely Deepfake Detection (Section 2) Deepfake Attribution And Recognition (Section 3), Passive Deepfake Authentication Methods (Section 4), Deepfakes Detection Method On Realistic Scenarios (Section 5.3), and Active Authentication (Section 6.3).
In this context, authors of the proposed papers are involved in the FF4ALL initiative (FF4ALL - Detection of Deep Fake Media and Life-Long Media Authentication), which aims to develop theoretical and practical tools for detecting and combating media counterfeits or Deepfakes, tracing their origin and limiting their dissemination. In the following sections, a brief overview of the main algorithms that aim to address the above-mentioned challenges will be presented.
## 2 Deepfake Detection
Deepfake technology poses significant challenges due to its potential for misuse, which can severely impact public well-being and trust. While current detection methods, primarily based on convolutional neural networks and deep learning paradigms, have shown promising results, they often struggle to generalize across the varied techniques employed in digital content manipulation. This issue primarily arises from the intricate interplay between textures and artifacts in Deepfake data, which traditional detection methods frequently overlook. In this context, artifacts are unintentional distortions or irregularities that occur during the Deepfake generation process, including unusual pixel formations or edge anomalies. Conversely, textures refer to the inherent patterns and fine details present in authentic images, such as the natural appearance of skin and hair. In fact, it is well established that synthetic manipulations typically disrupt the texture consistency of original images [77] and often leave detectable traces in the form of artifacts in both spatial [18] and frequency domains [23], particularly in specific facial regions [81]. Consequently, numerous studies focus their analysis on specific portions of face images to identify these inconsistencies. One promising approach involves using both No-Reference (NR) and Full-Reference (FR) quality measures to detect subtle manipulations in video frames [21]. This method has significantly improved cross-manipulation generalization by focusing on areas susceptible to artifacts, such as the mouth and eyes, and analyzing
the image quality degradation caused by Deepfake algorithms. In addition to artifacts, texture analysis provides another robust basis for distinguishing between real and fake images. In some Deepfake technologies dedicated to faceswapping operations, the inner and outer faces have different identities, making texture inconsistencies particularly evident [48]. However, focusing exclusively on either artifacts or textures in Deepfake detection can be limiting. While these approaches yield high accuracy in specific contexts, they often fail to adapt to new and evolving Deepfake techniques. To address this limitation, a novel framework called the Texture and Artifact Detector (TAD) has been proposed [26]. The TAD framework enhances Deepfake detection by leveraging both texture and artifact inconsistencies, thereby improving model generalization across various forgery scenarios through ensemble learning. Unfortunately, the performance of these methods is often hindered by the challenges posed by highly compressed data. High compression ratios can obscure subtle manipulations, leading to a significant degradation in detection accuracy [41]. A promising solution involves leveraging a learnable adaptive high-frequency enhancement framework to enrich weak high-frequency details in compressed content, thereby enhancing the robustness of Deepfake detection under compression [27]. Further details on compression impacts and related detection strategies will be discussed in subsequent sections.
## 3 Deepfake Attribution and Recognition
## 3.1 Deepfake Fingerprint and Attribution
Deepfake attribution, often referred to as Deepfake Model Recognition [37,69], encompasses methodologies capable of identifying the specific model used to generate synthetic data. This process includes attempts to estimate the unique model weights [5] of the architecture instance responsible for creating the Deepfake. SOTA techniques are highly effective in detecting Deepfake content generated by widely-used GANs [32,33,28] and DMs [35,68]. These techniques can even specialize to recognize the specific architectures, and, in more details, the specific model used in the creation procedure. Then, a more advanced challenge in this domain is identifying the exact model instance, characterized by a unique set of weights and parameters, within a given architecture: Guarnera et al. [37] demonstrated that using a simple ResNET-18 [38] engine combined with a metric learning approach [53], excellent results can be achieved in identifying the specific model used for creating synthetic data from 100 different instances of StyleGAN2-ADA [44]. A robust model recognition solution would enable the attribution of an image to a specific model owner, which is crucial for intellectual property rights [49]. To establish the ownership or authenticity of an image generated by a particular model within a specific architecture, new strategies and appropriate metrics are required [43]. In the context of forensic investigations involving Deepfake images, videos, or audio, state-of-the-art Deepfake detectors and architecture classifiers can be likened to the task of identifying camera models in traditional forensic analysis. Deepfake model recognition aims to trace the
origin of a Deepfake to a specific model instance within an architecture. This parallel underscores the necessity of developing advanced techniques for Deepfake model attribution to ensure authenticity in digital media.
## 4 Passive Deepfake Authentication Methods
In the modern era, where video calls have become a cornerstone of global communication, the importance of authenticating audio and video streams cannot be overstated. The advent of Deepfake technology poses a significant challenge to the integrity of digital communication. Traditional Deepfake detection methods may fall short as they often focus on either audio or video data in isolation. However, Deepfakes may involve sophisticated manipulations of both audio and video streams, making them harder to detect with monomodal methods.
This highlights the need for a multimodal approach that simultaneously analyzes both audio and visual data [73]. By correlating information from these two channels, we can significantly improve the accuracy of Deepfake detection. This approach takes advantage of the fact that inconsistencies are often more noticeable when multiple modalities of data are considered together. For instance, [40] leverages the incongruity between emotional cues portrayed by audio and visual modalities. In [4], the authenticity of a speaker is verified by detecting anomalous correspondences between his/her facial movements and what he/she says. Moreover, the results of [46] show that an ensemble of audio and visual baselines outperforms monomodal counterparts.
Given that audio-visual authentication methods may exploit monomodal detectors in a synergistic fashion, another step towards better performance is to enhance monomodal audio or visual detectors separately through the use of advanced techniques. Concerning the visual component, it is possible to leverage fusion of multiple detectors trained on purpose to capture orthogonal traces [60]. Concerning audio, it is possible to exploit modern solutions such as transformers [76], as well as investigating the use of semantic traces [7]. Despite the great effort of the multimedia forensics community, a series of challenges remains. Concerning multimodal solutions, the need for audio-video Deepfake datasets is becoming more than an urgent necessity. Indeed, most of the effort has been put towards monomodal datasets creation. Moreover, given the trend of large language models, it could be interesting to try using the same logic for audio visual reasoning. Concerning monomodal solutions, explainability has definitely not been reach yet, which still proves a problem in case of court of laws.
## 5 Deepfakes Detection Method on Realistic Scenarios
## 5.1 Deepfake Detection of image-videos in the wild
In recent years, there has been a growing interest in the study of techniques for the detection of Deepfake and AI-generated media [65,57]. Consequently, numerous solutions have been proposed to address the problems posed by the increasing spread of fake multimedia content. However, most of these solutions
perform well only in controlled settings, such as laboratory experiments, but fail to provide reliable results in real-life conditions typical of practical applications. Deep Learning (DL) models can effectively detect Deepfake media and identify their source. However, despite promising results, DL-based methods face several significant challenges, particularly in real-world applications where controlled laboratory conditions are absent. Firstly, DL models require vast amounts of labeled data for training, which is often difficult to obtain in real-life scenarios. Additionally, these models must handle unforeseen situations that were not accounted for during training, a common issue in multimedia forensics applied outside the lab. For DL models to be effective in the ongoing battle between forensic analysts and counterfeiters, it is crucial to address the risk of overfitting to training data, which can lead to failures in new, unexpected situations. While it is possible to train highly accurate detectors, these methods struggle to generalize to new generative techniques due to data drift [64]. Detectors perform well on the techniques they were trained on but often fail with content from new generative models. Another major obstacle is the black-box nature of DL techniques, making it difficult to interpret analysis results and understand decision-making processes. This lack of transparency hampers the practical application of DL in scenarios where accountability is essential. To address these issues, researchers address their attention on several strategies, including the use of one-class classifiers trained only on pristine images [47,1], developing classifiers with rejection options to opt out when encountering unfamiliar inputs not well-represented in the training set [58,2] and adopting methods capable of generalization though features fusion [50] together with multimodal approaches that combine audio and video streams [63,85]. Even considering these efforts, the practical application of automatic detectors has been minimal. Deploying these tools in commercial or mass verification systems presents numerous challenges beyond generalizing from a few known benchmarks [72,22,51,87]. One significant challenge is the need to continuously train these models on new generative/Deepfake techniques in a continual learning fashion [79]. Continual learning, also known as lifelong or incremental learning, is an ongoing approach to maintaining good model performance on evolving tasks without experiencing catastrophic forgetting [25]. This approach is well-suited to recognize content generated by new techniques and continuously adjusting models to account for data drift observed during inference versus training. A promising direction is the creation of an end-to-end Deepfake detection system that supports continuous integration and continuous delivery/deployment (CI/CD) with the design of a Machine Learning Model Operations ( MLOps [64,74]) pipeline, enabling the end-to-end development of continuously trained and monitored intelligent detectors with a minimal set of components.
## 5.2 Deepfake and Social Media
An extremely challenging 'real-world" case scenario involves the detection of Deepfake multimedia shared on social networks [11,62,67]. In fact, to cope with
bandwidth and storage limitations [11], social networks apply severe data compression and resizing. However, such processing, while reducing the overall multimedia size, also reduces the presence of forensic features used for the discrimination of real vs. fake multimedia [11,83,59]. A first study on GAN images [62] shared on Twitter demonstrates the adversarial effects of social network compression on Deepfake detectors. Specifically, while the visual quality of the shared images was untouched, the presence of forensic traces and patterns was reduced. The effects of social network sharing were then extensively observed and studied in [11]. In [67], the authors explore the challenges and advancements in the field of media forensics as applied to social network. The study addresses the increasing concerns regarding the authenticity and reliability of digital media shared on social platforms, focusing on challenges affecting source attribution algorithms [55] as well as multimedia verification [83]. For the latter, additional effort was devoted to studying whether the multimedia content is consistent with its descriptive text. The paper [67] discusses the emerging challenges in the field, such as the rise of Deepfakes and the use of bots for spreading disinformation. The authors highlight the need for advanced forensic tools to keep up with these sophisticated methods of media manipulation. In [11], the authors composed a large and diverse dataset counting 80k fake images generated with StyleGAN models and 70k real images collected from several state-of-the-art datasets [45]. In addition to this, while revealing insightful details on the entity and severity of social network compression applied by Twitter, Facebook and Telegram, the authors shown how their dataset can be used to finetuning new detectors, preserving their accuracy on social network compressed images while without experiencing ' catastrophic forgetting loss " [25]. Interestingly, [8] showed that while social networks degrade the presence of forensic artifacts used by real vs. fake detectors, they introduce other traces that can be exploited to reconstruct the life cycle of multimedia and determine on which social networks it has been shared. While the life cycle of multimedia does not provide specific information on its nature (i.e., real or fake), it can be fundamental in recovering the version of that multimedia closer to the original, unshared one. This, in turn, allows for more accurate real vs. fake detection. Finally, preliminary studies have been conducted also on videos shared on social netowrks [61], showing similar effects to those on images. One of these works [61,42] studies the effects of social network compression on FaceForensics [61] videos shared on Facebook and Youtube. The study provides a results in line with what observed on images [61] and a new dataset of shared videos to be used to finetune real vs. fake detectors. Nevertheless, while new works on Deepfake detection on social networks are available, the continuous update of social network compression algorithms makes the arms race increasingly challenging. As a consequence, this require additional effort to develop new architectures and updated datasets of social network shared images.
## 5.3 Detection of Deepfake Images and Videos in Adversarial Setting
An additional problem affecting virtually all the Deepfake forensic techniques developed so far is that such techniques are thought to operate in a benign set-
ting, that is, by neglecting the possible efforts made by an adversary to mislead the forensic analysis. Yet, recent researches [78,29] have shown how easy is to generate adversarial contents capable of deceiving image and video processing techniques based on DL, when the adversary is informed about the details of the tools employed by the analyst. Some works [66,9] have also studied the transferability of adversarial examples to networks different than those targeted by the attack, opening the way to the development of powerful attacks even when the attacker is unaware, or only partially aware, of the techniques used by the analyst. Even worse, often it is not even necessary that the adversary applies sophisticated attacks relying on the full or partial knowledge of the to-be-attacked system. By relying on the lack of robustness and the generalization capabilities of the forensic tools already outlined in Section 5.1, the adversary may simply process the deepkfake content in such way to prevent a correct forensic analysis, or at least degrade its performance up to a point to make it unusable. Some examples of this kind of attacks, often referred to as laundering attacks , include the application of moderate to strong lossy compression, geometric processing of images and videos, noise addition, histogram stretching and many others.
Understanding and ensuring the security of Deepfake forensic tools is a crucial problem, if such tools have to be used under the intrinsically adversarial conditions typical of multimedia forensics applications. For this reason, several efforts have been made to defend against adversarial attacks [14,52], both in the realm of computer vision applications and multimedia forensics. Still, no general effective solutions have been found yet [6]. Among the solutions developed so far, adversarial training [56] has received some consensus and has proven to at least mitigate the effectiveness of adversarial attacks in computer vision applications. As argued in [82], adversarial training forces DL models to focus on robust, possibly semantic, features, which are inherently more difficult to attack. Whether such a beneficial effect of adversarial training also applies to Deepfake forensic applications is still an open problem. It is no clear, in fact, if in multimedia forensics the equivalent of semantic computer vision features exist or not.
With regard to laundering attacks, the solutions proposed so far are similar to those already discussed in Section 5.1, given the, ultimately, the effectiveness of laundering attacks can be drastically reduced by improving the robustness and generalization capabilities of the forensic tools. A common approach to do so, involves the use of data augmentation techniques that enrich the training set with processed samples, thus improving the robustness of the forensic tools against the processing operators included in the data augmentation procedure. Yet, accounting for all possible kinds of processing during training is clearly unfeasible. Among the solutions proposed so far, we mention the possibility of identifying a kind of worst possible laundering attack and include it in the training procedure. Examples of such an approach are described in [10,70]. Despite all the efforts made, even for laundering attacks, a definitive solution has not been devised yet, thus adding yet another point to the to-do list of multimedia forensic researchers.
## 6 Active Authentication
## 6.1 Active deepefake detection
Passive Deepfake detection techniques [20,19] work a posteriori, after that the forged content has been generated, distributed and possibly processed, on the contrary, active methods work in a preemptive way, pre-processing the media in such a way to ease the subsequent analysis. This is the case, for instance, of Deepfake detection methods based on DNN watermarking, whereby the content generated by DNNs is watermarked in such a way to ease the distinction between genuine and fake media, and the attribution of the fake content to the network which generated it. An alternative possibility is to modify the computational imaging chain characterizing modern acquisition devices, to insert within the generated content a unique fingerprint to be used later on for authentication purposes. This means a change of paradigm that needs to be properly explored. Active authentication techniques represent a valid, and more reliable, alternative (or complement) to passive authentication, whenever the operating conditions allow their use.
Watermarking has recently been proposed as a means to protect the IPR of DNNs [24]. By tying a watermark to a DNN model, in fact, it would be possible to prove the ownership of the model or trace its illegal use. On this basis, DNN watermarking can be also used to link AI-generated contents [86], like Deepfakes, to the model which generated them, thus providing an easy and convincing way to distinguish between synthetic and natural contents. Such a goal is achieved by requiring that all the contents generated by a network contain a predefined watermark (a kind of synthetic fingerprint) that can be used later on to distinguish the synthetic images (or videos) generated by a trained model from real ones. This marks a drastic paradigm change with respect to current solutions based on multimedia forensics, since authentication is now achieved with the active help of the party which trained the media-generation network. Though some solutions have appeared in this direction, putting this idea at work requires that considerable advances are made particularly in terms of watermark robustness against image, video and network manipulations, security against adversarial attacks, payload and also imperceptibility. Succeeding into designing a new class of robust and secure solutions, based on active approaches, for Deepfake detection/attribution surely represents an open challenge and an interesting opportunity for scientific research in the field of multimedia forensics.
## 6.2 Efficient Media Origin Authentication
Customary deep-fake detection methods, both passive and active, are subject to false positives and false negatives, whose rate highly depends on the employed method and the goodness of the training data. False positives are due to various factors, such as the complexity of the content, the quality of the training data, or the intrinsic limitations of the detection algorithm itself. False negatives could happen if the deep-fake is very well made, or in general if the detection
method fails to recognize certain patterns or features that indicate a deep-fake. On the other hand, cryptographic signatures are 'almost perfect' from this point of view, in the sense that false negatives (i.e., authentic signatures which are not recognized so) are zero and false positives (i.e., fake signatures that are taken as authentic) are considered computationally infeasible to forge. This suggests that cryptographic signatures could be used fruitfully to detect deep-fakes (or better, the absence of deep-fakes) with perfect precision. In this direction, the work-in-progress standard JPEG Trust [80] by the Joint Photographic Experts Group (JPEG) aims to establish trust in digital media by addressing aspects of authenticity, provenance, and integrity. JPEG Trust will provide a framework for establishing trust in media through secure annotation of media assets throughout their life cycle, using cryptography as a key component. Cryptographically signing media allows deep-fakes to be repudiated by the interested person, since the signature on them will be absent or invalid. Such an anti-fake signature should allow 'good' manipulation of original file (at least cropping), disallow 'bad' manipulation, but also be space efficient, to save bandwidth on web servers once the media file is disseminated. Unfortunately, customary signature schemes like ECDSA do not have these properties.
To address this challenge, a solution could rely on novel aggregatable signatures, such as the Boneh-Lynn-Shacham (BLS) signature [13,12], which has been successfully used in blockchain technologies like Ethereum 2.0 to optimize storage . The BLS signature scheme makes use of a novel form of cryptography 1 called pairing-based cryptography, which allows for a plethora of new functionalities like attribute-based encryption[31,71,75]. Aggregatable signatures could be employed in JPEG (possibly within the JPEG Trust standard itself) in such a way to permit benign alterations of the image like cropping while preventing malicious tampering, without increasing too much the bandwidth occupation on web servers.
## 6.3 Trusted Remote Media Processing on Cloud and Edge Computing Systems
In the emerging Smart Cities context, systems based on IoT (Internet of Things) play an important role to allow citizens to interact with the environment and to benefit from advanced services, such as video surveillance, intelligent traffic lightning, and air quality sensing. From a technological point of view, using sensors and actuators to automate services is strategic, but managing, configuring, and optimizing the digital infrastructures to adapt their behavior to the specific needs of the context is a big challenge, both in terms of system design and security. Deepefake media detection in these scenarios represent a challenge due to the nature of possible manipulations of future citizens day life, hence a more holistic approach should be considered where the media production occur, since
1 https://github.com/ethereum/consensus-specs/blob/dev/specs/phase0/beaconchain.md#bls-signatures
at the Edge Computing Systems [16]. In just 20 years, with the objective to increase system response and reduce communication latency, computation moved from mainframes and computing rooms towards Cloud Computing, Fog Computing, and lastly, Edge Computing. A Federated Cloud-Edge infrastructure is considered, where different administrative domains are in place and where Machine Learning software artifacts, in the assertion of Federated Learning even at the Edge, help to distribute intelligence in this scenario.
Multimedia acquisition devices based on IoT generate an unprecedented amount of data, with the need of developing Cloud-based video big data analytics frameworks. A distributed approach in video recording and elaboration systems, such as video surveillance systems based on IP cameras, is highly recommended to overcome the maximum storage or throughput limitation of Network Video Recorders installed on single machines. To perform such a variety of tasks, and to be able to modify a device's behavior on-demand, the Function as a Service (FaaS) computational paradigm is generally adopted. FaaS allows to define several minimal applications and to run one or more instances of these on the same device at the same time. FaaS framework relies on two configuration approaches: a local configuration file, generally YAML, or a secure remote server. However, both come with limitations: a local file configuration requires direct access to the device, physically or through a secure connection, to modify it. Alternatively, a remote server can store and send updated configuration files, but it might be vulnerable to well-known cyberattacks suck as Man-in-the-middle (MITM) or Distributed Denial of Service (DDoS), making it unusable and unreliable. To overcome such limitations, it is possible to benefit from three technologies that have been increasingly recognized to be able to address information access problems and system trustiness in different application domains: Federated Learning [3], Blockchain and IPFS (InterPlanetary File System) starting at the Edge:
- -Federated Learning is a decentralized approach to training Machine Learning Models. In traditional Machine Learning, data is centralized in the Cloud, where a single model is trained on the entire dataset. Federated Learning, on the other hand, allows for training Machine Learning Models across multiple decentralized devices or servers that hold local data samples without exchanging them. Moreover, Federated Learning at the Edge refers to the application of Federated Learning techniques on Edge devices, such as IoT devices, or Edge Servers. This approach combines the benefits of Federated Learning, which ensures Data Privacy and reduces communication costs, with the advantages of Edge Computing Systems, which enables data processing and model training to occur closer to where the data is generated, hence, fake Media might not exit from the Edge.
- -The use of Blockchain, supported by the flexibility and robustness of Smart Contracts, allows the combining of the well-known FaaS paradigm with the intrinsic features of data non-repudiation and immutability, replacing the service configuration with a Smart Contract, guaranteeing protection against distributed cyber-attacks [17].
- -IPFS is a distributed system for storing and accessing files. Since the block size of the Blockchain does not allow storing files, these can be uploaded to this special file storage, which produces a unique hash value to be used as a key to access its content [54].
## 7 Conclusion
In this paper, we have conducted a comprehensive review of the state-of-the-art techniques and challenges in Deepfake media forensics. Our exploration covered the core areas of Deepfake detection, attribution and recognition, passive authentication, detection in realistic scenarios, and active authentication. Each of these areas addresses specific facets of the Deepfake phenomenon, from the identification of synthetic media and tracing their origins to ensuring the robustness of detection systems in real-world environments and embedding verifiable information within media for instant authentication. Future work will focus on conducting a more in-depth analysis of practical countermeasures and gaining deeper insights into real-world applications (e.g. highly compressed data).
Acknowledgments. This study has been partially supported by SERICS (PE00000014) under the MUR National Recovery and Resilience Plan funded by the European Union - NextGenerationEU.
Disclosure of Interests. The authors have no competing interests.
## References
- 1. Abady, L., Dimitri, G.M., Barni, M.: A one-class classifier for the detection of gan manipulated multi-spectral satellite images. Remote Sensing 16 (5) (2024)
- 2. Abady, L., Wang, J., Tondi, B., Barni, M.: A siamese-based verification system for open-set architecture attribution of synthetic images. Pattern Recognition Letters 180 , 75-81 (2024)
- 3. Abdelgaber, Y.E., Ahmed, Y.A., Salem, M.A.M., Salem, M.A.G.: Federated learning for resource management in edge computing. In: 2023 Eleventh International Conference on Intelligent Computing and Information Systems (ICICIS). pp. 102109 (2023). https://doi.org/10.1109/ICICIS58388.2023.10413933
- 4. Agarwal, S., Hu, L., Ng, E., Darrell, T., Li, H., Rohrbach, A.: Watch those words: Video falsification detection using word-conditioned facial motion. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (2023)
- 5. Asnani, V., Yin, X., Hassner, T., Liu, X.: Reverse engineering of generative models: Inferring model hyperparameters from generated images. IEEE Transactions on Pattern Analysis and Machine Intelligence (2023)
- 6. Athalye, A., Carlini, N., Wagner, D.: Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In: Proceedings of PMLR, International conference on machine learning. pp. 274-283 (2018)
- 7. Attorresi, L., Salvi, D., Borrelli, C., Bestagini, P., Tubaro, S.: Combining automatic speaker verification and prosody analysis for synthetic speech detection. In: International Conference on Pattern Recognition (ICPR) (2022)
- 8. Baracchi, D., Boato, G., De Natale, F., Iuliani, M., Montibeller, A., Pasquini, C., Piva, A., Shullani, D.: Towards open-world multimedia forensics through media signature encoding. IEEE Access (2024)
- 9. Barni, M., andf E. Nowroozi, K.K., Tondi, B.: On the transferability of adversarial examples against CNN-based image forensics. In: Proceedings ICASSP - IEEE International Conference on Acoustics, Speech and Signal Processing. pp. 82868290 (2019)
- 10. Barni, M., Nowroozi, E., Tondi, B.: Higher-order, adversary-aware, double JPEGdetection via selected training on attacked samples. In: Proceedings of 25th European signal processing conference (EUSIPCO). pp. 281-285 (2017)
- 11. Boato, G., Pasquini, C., Stefani, A.L., Verde, S., Miorandi, D.: Trueface: a dataset for the detection of synthetic face images from social networks. In: 2022 IEEE International Joint Conference on Biometrics (IJCB). pp. 1-7 (2022). https://doi.org/10.1109/IJCB54206.2022.10007988
- 12. Boneh, D., Gorbunov, S., Wahby, R.S., Wee, H., Wood, C.A., Zhang, Z.: BLS Signatures. Internet-Draft draft-irtf-cfrg-bls-signature-05, Internet Engineering Task Force (Jun 2022), work in Progress
- 13. Boneh, D., Lynn, B., Shacham, H.: Short signatures from the weil pairing. In: Boyd, C. (ed.) Advances in Cryptology - ASIACRYPT 2001. pp. 514-532. Springer Berlin Heidelberg, Berlin, Heidelberg (2001)
- 14. Bountakas, P., Zarras, A., Lekidis, A., Xenakis, C.: Defense strategies for adversarial machine learning: A survey. Computer Science Review (2023)
- 15. Casu, M., Guarnera, L., Caponnetto, P., Battiato, S.: Genai mirage: The impostor bias and the deepfake detection challenge in the era of artificial illusions. Forensic Science International: Digital Investigation 50 , 301795 (2024). https://doi.org/https://doi.org/10.1016/j.fsidi.2024.301795
- 16. Catalfamo, A., Celesti, A., Fazio, M., Randazzo, G., Villari, M.: A platform for federated learning on the edge: a video analysis use case. In: 2022 IEEE Symposium on Computers and Communications (ISCC). pp. 1-7 (2022). https://doi.org/10.1109/ISCC55528.2022.9912968
- 17. Catalfamo, A., Ruggeri, A., Celesti, A., Fazio, M., Villari, M.: A microservices and blockchain based one time password (mbb-otp) protocol for security-enhanced authentication. In: 2021 IEEE Symposium on Computers and Communications (ISCC). pp. 1-6 (2021). https://doi.org/10.1109/ISCC53001.2021.9631479
- 18. Chai, L., Bau, D., Lim, S.N., Isola, P.: What makes fake images detectable? understanding properties that generalize. In: Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16. pp. 103-120. Springer (2020)
- 19. Ciamarra, A., Caldelli, R., Becattini, F., Seidenari, L., Del Bimbo, A.: Deepfake detection by exploiting surface anomalies: The surfake approach. In: 2024 IEEE/CVF Winter Conference on Applications of Computer Vision Workshops (WACVW). pp. 1024-1033 (2024). https://doi.org/10.1109/WACVW60836.2024.00112
- 20. Coccomini, D.A., Zilos, G.K., Amato, G., Caldelli, R., Falchi, F., Papadopoulos, S., Gennaro, C.: Mintime: Multi-identity size-invariant video deepfake detection. IEEE Transactions on Information Forensics and Security pp. 1-1 (2024). https://doi.org/10.1109/TIFS.2024.3409054
- 21. Concas, S., La Cava, S.M., Casula, R., Orrù, G., Puglisi, G., Marcialis, G.L.: Quality-based artifact modeling for facial deepfake detection in videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3845-3854 (2024)
- 22. Corvi, R., Cozzolino, D., Poggi, G., Nagano, K., Verdoliva, L.: Intriguing properties of synthetic images: from generative adversarial networks to diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 973-982 (2023)
- 23. Durall, R., Keuper, M., Keuper, J.: Watch your up-convolution: Cnn based generative deep neural networks are failing to reproduce spectral distributions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 7890-7899 (2020)
- 24. Fei, J., Xia, Z., Tondi, B., Barni, M.: Supervised gan watermarking for intellectual property protection. In: 2022 IEEE International Workshop on Information Forensics and Security (WIFS). pp. 1-6 (2022)
- 25. French, R.M.: Catastrophic forgetting in connectionist networks. Trends in Cognitive Sciences 3 (4), 128 - 135 (1999)
- 26. Gao, J., Micheletto, M., Orrù, G., Concas, S., Feng, X., Marcialis, G.L., Roli, F.: Texture and artifact decomposition for improving generalization in deep-learningbased deepfake detection. Engineering Applications of Artificial Intelligence 133 , 108450 (2024)
- 27. Gao, J., Xia, Z., Marcialis, G.L., Dang, C., Dai, J., Feng, X.: Deepfake detection based on high-frequency enhancement network for highly compressed content. Expert Systems with Applications 249 , 123732 (2024)
- 28. Giudice, O., Guarnera, L., Battiato, S.: Fighting deepfakes by detecting gan dct anomalies. Journal of Imaging 7 (8), 128 (2021)
- 29. Goodfellow, I.J., Shlens, J., Szegedy, C.: Explaining and harnessing adversarial examples. arXiv:1412.6572 (2014)
- 30. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative Adversarial Nets. Advances in Neural Information Processing Systems 27 (2014)
- 31. Goyal, V., Pandey, O., Sahai, A., Waters, B.: Attribute-based encryption for fine-grained access control of encrypted data. In: Proceedings of the 13th ACM Conference on Computer and Communications Security. p. 89-98. CCS '06, Association for Computing Machinery, New York, NY, USA (2006). https://doi.org/10.1145/1180405.1180418
- 32. Guarnera, L., Giudice, O., Battiato, S.: Deepfake detection by analyzing convolutional traces. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 666-667 (2020)
- 33. Guarnera, L., Giudice, O., Battiato, S.: Fighting deepfake by exposing the convolutional traces on images. IEEE Access 8 , 165085-165098 (2020)
- 34. Guarnera, L., Giudice, O., Battiato, S.: Level up the deepfake detection: a method to effectively discriminate images generated by gan architectures and diffusion models. arXiv preprint arXiv:2303.00608 (2023)
- 35. Guarnera, L., Giudice, O., Battiato, S.: Mastering deepfake detection: A cuttingedge approach to distinguish gan and diffusion-model images. ACM Transactions on Multimedia Computing, Communications and Applications (2024)
- 36. Guarnera, L., Giudice, O., Nastasi, C., Battiato, S.: Preliminary forensics analysis of deepfake images. In: 2020 AEIT international annual conference (AEIT). pp. 16. IEEE (2020)
- 37. Guarnera, L., Giudice, O., Nießner, M., Battiato, S.: On the exploitation of deepfake model recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 61-70 (2022)
- 38. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 770-778 (2016)
- 39. Ho, J., Jain, A., Abbeel, P.: Denoising Diffusion Probabilistic Models. Advances in Neural Information Processing Systems 33 , 6840-6851 (2020)
- 40. Hosler, B., Salvi, D., Murray, A., Antonacci, F., Bestagini, P., Tubaro, S., Stamm, M.C.: Do deepfakes feel emotions? a semantic approach to detecting deepfakes via emotional inconsistencies. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (2021)
- 41. Hu, J., Liao, X., Wang, W., Qin, Z.: Detecting compressed deepfake videos in social networks using frame-temporality two-stream convolutional network. IEEE Transactions on Circuits and Systems for Video Technology 32 (3), 1089-1102 (2021)
- 42. Hu, J., Liao, X., Wang, W., Qin, Z.: Detecting compressed deepfake videos in social networks using frame-temporality two-stream convolutional network. IEEE Transactions on Circuits and Systems for Video Technology 32 (3), 1089-1102 (2022). https://doi.org/10.1109/TCSVT.2021.3074259
- 43. Huang, Z., Li, B., Cai, Y., Wang, R., Guo, S., Fang, L., Chen, J., Wang, L.: What can discriminator do? towards box-free ownership verification of generative adversarial networks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 5009-5019 (2023)
- 44. Karras, T., Aittala, M., Hellsten, J., Laine, S., Lehtinen, J., Aila, T.: Training generative adversarial networks with limited data. In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M.F., Lin, H. (eds.) Advances in Neural Information Processing Systems. vol. 33, pp. 12104-12114. Curran Associates, Inc. (2020)
- 45. Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401-4410 (2019)
- 46. Khalid, H., Kim, M., Tariq, S., Woo, S.S.: Evaluation of an audio-video multimodal deepfake dataset using unimodal and multimodal detectors. In: Workshop on synthetic multimedia-audiovisual deepfake generation and detection (2021)
- 47. Khalid, H., Woo, S.S.: Oc-fakedect: Classifying deepfakes using one-class variational autoencoder. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). pp. 2794-2803 (2020)
- 48. La Cava, S.M., Orrù, G., Drahansky, M., Marcialis, G.L., Roli, F.: 3d face reconstruction: The road to forensics. ACM Computing Surveys 56 (3), 1-38 (2023)
- 49. Leotta, R., Giudice, O., Guarnera, L., Battiato, S.: Not with my name! inferring artists' names of input strings employed by diffusion models. In: International Conference on Image Analysis and Processing. pp. 364-375. Springer (2023)
- 50. Leporoni, G., Maiano, L., Papa, L., Amerini, I.: A guided-based approach for deepfake detection: Rgb-depth integration via features fusion. Pattern Recognition Letters 181 , 99-105 (2024). https://doi.org/https://doi.org/10.1016/j.patrec.2024.03.025
- 51. Li, C., Huang, Z., Paudel, D.P., Wang, Y., Shahbazi, M., Hong, X., Van Gool, L.: A continual deepfake detection benchmark: Dataset, methods, and essentials. p. 1339 - 1349 (2023)
- 52. Liang, H., He, E., Zhao, Y., Jia, Z., Li, H.: Adversarial attack and defense: A survey. Electronics (2022)
- 53. Liu, E.Y., Guo, Z., Zhang, X., Jojic, V., Wang, W.: Metric learning from relative comparisons by minimizing squared residual. In: 2012 IEEE 12th International Conference on Data Mining. pp. 978-983. IEEE (2012)
- 54. Lukaj, V., Martella, F., Fazio, M., Galletta, A., Celesti, A., Villari, M.: Gatewaybased certification approach to include iot nodes in a trusted edge/cloud environment. In: 2023 IEEE/ACM 23rd International Symposium on Cluster, Cloud and Internet Computing Workshops (CCGridW). pp. 237-241 (2023)
- 55. Lukáš, J., Fridrich, J., Goljan, M.: Detecting digital image forgeries using sensor pattern noise. In: Security, steganography, and watermarking of multimedia contents VIII. vol. 6072, pp. 362-372. SPIE (2006)
- 56. Madry, A., Makelov, A., Schmidt, L., Tsipras, D., Vladu, A.: Towards deep learning models resistant to adversarial attacks. arXiv:1706.06083 (2017)
- 57. Maiano, L., Benova, A., Papa, L., Stockner, M., Marchetti, M., Convertino, G., Mazzoni, G., Amerini, I.: Human versus machine: A comparative analysis in detecting artificial intelligence-generated images. IEEE Security &; Privacy 22 (03), 77-86 (may 2024). https://doi.org/10.1109/MSEC.2024.3390555
- 58. Maier, A., Riess, C.: Reliable out-of-distribution recognition of synthetic images. Journal of Imaging 10 (5), 110 (2024)
- 59. Maier, A., Riess, C.: Reliable out-of-distribution recognition of synthetic images. Journal of Imaging 10 (5), 110 (2024)
- 60. Mandelli, S., Bonettini, N., Bestagini, P., Tubaro, S.: Detecting gan-generated images by orthogonal training of multiple cnns. In: IEEE International Conference on Image Processing (ICIP). pp. 3091-3095 (2022)
- 61. Marcon, F., Pasquini, C., Boato, G.: Detection of manipulated face videos over social networks: A large-scale study. Journal of Imaging 7 (10), 193 (2021)
- 62. Marra, F., Gragnaniello, D., Cozzolino, D., Verdoliva, L.: Detection of gangenerated fake images over social networks. In: 2018 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR). pp. 384-389 (2018)
- 63. Mongelli, L., Maiano, L., Amerini, I.: CMDD: A novel multimodal two-stream CNN deepfakes detector. In: Petrocchi, M., Viviani, M. (eds.) Proceedings of the 4th Workshop on Reducing Online Misinformation through Credible Information Retrieval co-located with the 46th European Conference on Information Retrieval, ROMCIR@ECIR 2024, Glasgow, UK, March 24, 2024. CEUR Workshop Proceedings, vol. 3677, pp. 17-30. CEUR-WS.org (2024)
- 64. Paleyes, A., Urma, R.G., Lawrence, N.D.: Challenges in deploying machine learning: A survey of case studies. ACM Computing Surveys 55 (6) (2022)
- 65. Papa, L., Faiella, L., Corvitto, L., Maiano, L., Amerini, I.: On the use of stable diffusion for creating realistic faces: from generation to detection. In: 11th International Workshop on Biometrics and Forensics, IWBF 2023, Barcelona, Spain, April 19-20, 2023. pp. 1-6. IEEE (2023). https://doi.org/10.1109/IWBF57495.2023.10156981
- 66. Papernot, N., McDaniel, P., Goodfellow, I., Jha, S., Celik, Z.B., Swami, A.: Practical black-box attacks against machine learning. In: Proceedings of the 2017 ACM on Asia conference on computer and communications security. pp. 506-519 (2017)
- 67. Pasquini, C., Amerini, I., Boato, G.: Media forensics on social media platforms: a survey. EURASIP Journal on Information Security 2021 (1), 4 (2021)
- 68. Pontorno, O., Guarnera, L., Battiato, S.: Deepfeaturex net: Deep features extractors based network for discriminating synthetic from real images. arXiv preprint arXiv:2404.15697 (2024)
- 69. Pontorno, O., Guarnera, L., Battiato, S.: On the exploitation of dct-traces in the generative-ai domain. arXiv preprint arXiv:2402.02209 (2024)
- 70. Purnekar, N., Abady, L., Tondi, B., Barni, M.: Improving the robustness of synthetic images detection by means of print and scan augmentation. In: Proceedings of Information Hiding & Multimedia Security Conference (IH&MMSEC) (2024)
- 71. Rasori, M., Perazzo, P., Dini, G.: ABE-Cities: An attribute-based encryption system for smart cities. In: 2018 IEEE International Conference on Smart Computing (SMARTCOMP). pp. 65-72 (2018)
- 72. Rossler, A., Cozzolino, D., Verdoliva, L., Riess, C., Thies, J., Niessner, M.: Faceforensics++: Learning to detect manipulated facial images. vol. 2019-October, p. 1 - 11 (2019)
- 73. Salvi, D., Liu, H., Mandelli, S., Bestagini, P., Zhou, W., Zhang, W., Tubaro, S.: A robust approach to multimodal deepfake detection. Journal of Imaging 9 (6) (2023). https://doi.org/10.3390/jimaging9060122
- 74. Semola, R., Lomonaco, V., Bacciu, D.: Continual-learning-as-a-service (claas): Ondemand efficient adaptation of predictive models. arXiv preprint arXiv:2206.06957 (2022)
- 75. Sicari, S., Rizzardi, A., Dini, G., Perazzo, P., La Manna, M., Coen-Porisini, A.: Attribute-based encryption and sticky policies for data access control in a smart home scenario: a comparison on networked smart object middleware. International Journal of Information Security 20 , 695-713 (2021)
- 76. Singh Yadav, A.K., Bhagtani, K., Baireddy, S., Bestagini, P., Tubaro, S., Delp, E.J.: Mdrt: Multi-domain synthetic speech localization. In: IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (2024)
- 77. Sun, X., Wu, B., Chen, W.: Identifying invariant texture violation for robust deepfake detection. arXiv preprint arXiv:2012.10580 (2020)
- 78. Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., Fergus, R.: Intriguing properties of neural networks. arXiv:1503.02531 (2013)
- 79. Tassone, F., Maiano, L., Amerini, I.: Continuous fake media detection: adapting deepfake detectors to new generative techniques (2024)
- 80. Temmermans, F., Caldwell, S., Papadopoulos, S., Pereira, F., Rixhon, P.: Towards an international standard to establish trust in media production, distribution and consumption. In: 2023 24th International Conference on Digital Signal Processing (DSP). pp. 1-5 (2023)
- 81. Tolosana, R., Romero-Tapiador, S., Vera-Rodriguez, R., Gonzalez-Sosa, E., Fierrez, J.: Deepfakes detection across generations: Analysis of facial regions, fusion, and performance evaluation. Engineering Applications of Artificial Intelligence 110 , 104673 (2022)
- 82. Tsipras, D., Santurkar, S., Engstrom, L., Turner, A.: Robustness may be at odds with accuracy. arXiv:1805.12152 (2018)
- 83. Verdoliva, L.: Media forensics and deepfakes: An overview. IEEE Journal on Selected Topics in Signal Processing 14 (5), 910 - 932 (2020)
- 84. Wang, S.Y., Wang, O., Zhang, R., Owens, A., Efros, A.A.: Cnn-generated images are surprisingly easy to spot... for now. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8695-8704 (2020)
- 85. Wani, T.M., Amerini, I.: Deepfakes audio detection leveraging audio spectrogram and convolutional neural networks. In: Image Analysis and Processing - ICIAP 2023 - 22nd International Conference, ICIAP 2023, Udine, Italy, September 11-15, 2023, Proceedings, Part II. Lecture Notes in Computer Science, vol. 14234, pp. 156-167. Springer (2023)
- 86. Yu, N., Skripniuk, V., Abdelnabi, S., Fritz, M.: Artificial fingerprinting for generative models: Rooting deepfake attribution in training data. In: 2021 IEEE/CVF International Conference on Computer Vision (ICCV). pp. 14428-14437 (2021). https://doi.org/10.1109/ICCV48922.2021.01418
- 87. Zi, B., Chang, M., Chen, J., Ma, X., Jiang, Y.G.: Wilddeepfake: A challenging real-world dataset for deepfake detection. p. 2382 - 2390 (2020) | null | [
"Irene Amerini",
"Mauro Barni",
"Sebastiano Battiato",
"Paolo Bestagini",
"Giulia Boato",
"Tania Sari Bonaventura",
"Vittoria Bruni",
"Roberto Caldelli",
"Francesco De Natale",
"Rocco De Nicola",
"Luca Guarnera",
"Sara Mandelli",
"Gian Luca Marcialis",
"Marco Micheletto",
"Andrea Montibeller",
"Giulia Orru'",
"Alessandro Ortis",
"Pericle Perazzo",
"Giovanni Puglisi",
"Davide Salvi",
"Stefano Tubaro",
"Claudia Melis Tonti",
"Massimo Villari",
"Domenico Vitulano"
] | 2024-08-01T08:57:47+00:00 | 2024-08-13T15:10:20+00:00 | [
"cs.CV"
] | Deepfake Media Forensics: State of the Art and Challenges Ahead | AI-generated synthetic media, also called Deepfakes, have significantly
influenced so many domains, from entertainment to cybersecurity. Generative
Adversarial Networks (GANs) and Diffusion Models (DMs) are the main frameworks
used to create Deepfakes, producing highly realistic yet fabricated content.
While these technologies open up new creative possibilities, they also bring
substantial ethical and security risks due to their potential misuse. The rise
of such advanced media has led to the development of a cognitive bias known as
Impostor Bias, where individuals doubt the authenticity of multimedia due to
the awareness of AI's capabilities. As a result, Deepfake detection has become
a vital area of research, focusing on identifying subtle inconsistencies and
artifacts with machine learning techniques, especially Convolutional Neural
Networks (CNNs). Research in forensic Deepfake technology encompasses five main
areas: detection, attribution and recognition, passive authentication,
detection in realistic scenarios, and active authentication. This paper reviews
the primary algorithms that address these challenges, examining their
advantages, limitations, and future prospects. |
Subsets and Splits