Add files using upload-large-folder tool

2401/2401.00001/main.tex

@@ -0,0 +1,289 @@
+\documentclass{article}
+\usepackage[utf8]{inputenc}
+\usepackage[letterpaper, portrait, margin=1in]{geometry}
+\usepackage{graphicx}
+\usepackage{neuralnetwork}
+\usepackage{siunitx}
+\usepackage{authblk}
+\DeclareUnicodeCharacter{FF0C}{,}
+\title{Sector Rotation by Factor Model and Fundamental Analysis}
+
+\author[1]{Runjia Yang}
+\affil[1]{University of California, Davis}
+\author[2]{Beining Shi}
+\affil[2]{University of California, Davis}
+
+\date{Sept 2023}
+
+\begin{document}
+
+\maketitle
+
+\section*{Abstract}
+This study presents an analytical approach to sector rotation, leveraging both factor models and fundamental metrics. We initiate with a systematic classification of sectors, followed by an empirical investigation into their returns. Through factor analysis, the paper underscores the significance of momentum and short-term reversion in dictating sectoral shifts. A subsequent in-depth fundamental analysis evaluates metrics such as PE, PB, EV-to-EBITDA, Dividend Yield, among others. Our primary contribution lies in developing a predictive framework based on these fundamental indicators. The constructed models, post rigorous training, exhibit noteworthy predictive capabilities. The findings furnish a nuanced understanding of sector rotation strategies, with implications for asset management and portfolio construction in the financial domain.
+
+\vspace{4mm}
+\noindent
+{\bf Keywords:} US Industrial Sectors, Factor Analysis, Fundamental Analysis, Trading Strategy.
+
+\section{Introduction}
+Sector is composed by a basket of stocks that representing companies in certain business class, which has unique features according to the business. Under certain conditions, such as economic cycles, sectors may behave accordingly due to the different characteristics of businesses. In this report, we are exploring how to capture returns by finding the hidden features behind different sectors and determining the leading sectors in some particular market conditions or social environments. Generally, this report covers a brief exploration of market and fundamental factors, explaining the meaning of each factors and how they are related to some sectors.Then we applied a neural network model to do a classification and prediction using the fundamental factors as inputs. At the end of the report, we also covers how sectors behaved under global events.
+
+\section{Sector Classification and Return Analysis}
+There are many different ways to divide sectors. For the purpose of common acceptance and convenience for future data acquirement, we used the MSCI Global Industry Classification Standard, which includes 11 level one sectors, 24 level two industry groups, 69 level three industries, and 158 sub-industries. We use the 11 level one sectors as our main target. They are Energy, Materials, Industrials, Consumer Discretionary, Consumer Staples, Health Care, Financials, Information Technology, Communication Services, Utilities, and Real Estate. In order to track the performance of each sector, we use the S$\&$P500 GICS Indices which are constructed exactly as the MSCI classification.Before working on any strategies further, we need to determine if there are actually possible profits. In our case, we need to check how big the differences between sectors' returns are. For each observation time period,
+{\textbf{Define:}}  Return Difference =($ \sum_{\text{top 3}}$ Sector Return - $\sum_{\text{bottom 3}}$  Sector Return )/3
+Based on a monthly frequency, we calculate the return difference and get the following plot.
+\begin{center}
+\includegraphics[scale=0.35]{return_difference.png}
+\end{center}
+Also, by calculation, the quarterly return difference has a mean of 0.1306, median of 0.1185, standard deviation of 0.0523. We can reach to a result that half of the quarterly return difference is more than 11.85$\%$. It is easy to see that there does exist potential investment opportunity by capturing the return differences between sectors. 
+
+\section{Factor Analysis}
+
+\subsection{Momentum Factor Exploration}
+Momentum premium was first recognized by UCLA scholars Narasimhan Jegadeesh and Sheridan Titman in 1993. The momentum premium is established on the observation that assets that have performed well in the past have the trend to persist good performance in the future. Though the momentum effect is considered to be a market anomaly, it has been recognized widely among many asset classes. We will explore the momentum effect based on the sector indices introduced above.
+
+\subsubsection{Factor Construction}
+First of all, we need to construct the momentum factor. Typically, the momentum factor is constructed by the past 6 or 12 months cumulative return and excludes the most recent month's return, considering that there are also short-term reversion effects based on the mean-reversion effects. However, without a clear idea of how the sector indices carry the momentum effect, we need to explore through time intervals to find the best possible momentum factor. Then we constructed 12 different momentum factors using the past 1 to 12 month's return and excluding the most recent 0.1 portion trading days of each time period to avoid short-term reversion. For each of the factors with the period of n months
+\begin{equation*}
+     MOM\_{nM} = \sum_{\text{Past} n*21 \text{ days}}  R_{d} - \sum_{\text{Past }  0.1*n*21 \text{ days}}   R_{d}
+\end{equation*}
+where $R_{d}$ is the daily return.
+\subsubsection{Calculate Factor Returns}
+For the 12 factors we got, we normalized them cross sections. Then we rank the factor exposures for each sector and take long positions of sectors with the highest two factor exposures, take short position of sectors with the lowest two factor exposures. Then we trade our portfolio under a monthly frequency. Here are the results from 2002 to 2022 February:
+\begin{center}
+\begin{tabular}{|l|S[table-format=3.1(3)]|S[table-format=3.1(3)]|}
+\hline
+{Factor}&{Factor Return}&{Sharpe Ratio} \\
+\hline
+{MOM\_1M} & -0.0297 & -0.09  \\\hline
+{MOM\_2M} & -0.0583 & -0.19  \\ \hline
+{MOM\_3M} & 0.0082 & 0.02  \\ \hline
+{MOM\_4M} & -0.0245 & -0.08  \\ \hline
+{MOM\_5M} & -0.0136 & -0.04  \\ \hline
+{MOM\_6M} & 0.0195 & 0.06  \\ \hline
+{MOM\_7M} & 0.0535 & 0.15  \\ \hline
+{MOM\_8M} & 0.0721 & 0.21  \\ \hline
+{MOM\_9M} & 0.0413 & 0.11  \\ \hline
+{MOM\_10M} & -0.0470 & -0.13  \\ \hline
+{MOM\_11M} & 0.0371 & 0.10  \\ \hline
+{MOM\_12M} & 0.0327 & 0.01 \\ \hline
+\end{tabular}
+\end{center}
+Since there are several market crashes where the momentum factor led to negative returns, we also take a look at the most recent five years from 2017 to 2022 February:
+\begin{center}
+\begin{tabular}{|l|S[table-format=3.1(3)]|S[table-format=3.1(3)]|}
+\hline
+{Factor}&{Factor Return}&{Sharpe Ratio} \\
+\hline
+{MOM\_1M} & 0.0213 & 0.06  \\\hline
+{MOM\_2M} & -0.0494 & -0.14  \\ \hline
+{MOM\_3M} & 0.1154 & 0.29  \\ \hline
+{MOM\_4M} & -0.1578 & 0.41  \\ \hline
+{MOM\_5M} & -0.1788 & 0.47  \\ \hline
+{MOM\_6M} & 0.1163 & 0.35  \\ \hline
+{MOM\_7M} & 0.2119 & 0.62  \\ \hline
+{MOM\_8M} & 0.1869 & 0.56  \\ \hline
+{MOM\_9M} & 0.1583 & 0.43  \\ \hline
+{MOM\_10M} & 0.0504 & 0.12  \\ \hline
+{MOM\_11M} & 0.1547 & 0.40  \\ \hline
+{MOM\_12M} & 0.1289 & 0.33 \\ \hline
+\end{tabular}
+\end{center}
+From this table, we can tell that by using the MOM$\_$7M factor, we can reach a maximum annual return rate of 21.19$\%$ and a maximum Sharpe ratio of 0.62. It is also interesting that we find the momentum factor with a short time period, for example, MOM$\_$1M and MOM$\_$2M, have a very small even negative return rate. However, it exactly conforms to the short term reversion effect that the typical momentum factor would exclude. 
+
+\subsection{Short Term Reversion Factor Exploration}
+Short term reversion factor follows the simple principle that asset's price will have the trend to stay on an average level. Since we can see from the previous results of the momentum factor that there does exist short term reversion effect, we can try different reversion factors and find out what would be the best short term reversion observation period.
+
+\subsubsection{Factor Construction}
+Similarly, we can define several reversion factors with different time periods. And we take the negative number of the past n days cumulative return as the factor exposures.
+\begin{equation*}
+    REV\_{nD} = - \sum_{\text{Past } n \text{ days}}  R_{d}
+\end{equation*}
+For the purpose of exploring the optimal time period, we take 5-day time interval and create 12 reversion factors from 5 trade days to 55 trade days.
+
+\subsubsection{Calculate Factor Returns}
+By using the same method, we compute the rank of each sector's factor exposure, and long the top two sectors, short the last two sectors on a monthly observation frequency. Between 2002 and 2022 February, the results are:
+    \begin{center}
+    \begin{tabular}{|l|S[table-format=3.1(3)]|S[table-format=3.1(3)]|}
+    \hline
+        {Factor} & {Factor Return} & {Sharpe Ratio}  \\ \hline
+        {Rev\_5D} & -0.0059 & -0.0597  \\ \hline
+        {Rev\_10D} & 0.0294 & 0.3320  \\ \hline
+        {Rev\_15D} & 0.0127 & 0.1338  \\ \hline
+        {Rev\_20D} & 0.0072 & 0.0715  \\ \hline
+        {Rev\_25D} & 0.0690 & 0.7768  \\ \hline
+        {Rev\_30D} & 0.0877 & 0.8735  \\ \hline
+        {Rev\_35D} & 0.0135 & 0.1288  \\ \hline
+        {Rev\_40D} & -0.0314 & -0.3422  \\ \hline
+        {Rev\_45D} & -0.0694 & -0.8417  \\ \hline
+        {Rev\_50D} & -0.0183 & -0.5919  \\ \hline
+        {Rev\_55D} & -0.0517 & -0.1863 \\ \hline
+    \end{tabular}
+    \end{center}
+
+From this table, we can tell that for the time between 2002 to recent time, the short term reversion effect is optimal for taking the past 30 days cumulative return. It has an optimal annual return rate of 8.77$\%$ on average and leads to a sharp ratio of 0.8735. 
+
+
+
+\section{Fundamental Analysis}
+Fundamental Analysis are always a good aspect to look at for investing. We collected quarterly data for all 11 indices from Bloomberg, including their P/E ratio, EV/EBIT, Profit Margin, etc. Our fundamental analysis would start from discovering features for each of the fundamental ratio,then we are trying to predict the sector performance by constructing using some of the features we found.
+
+\subsection{PE Ratio}
+The P/E is one of the most widely used tools to determine a stock’s relative valuation. The purpose of analyzing the ratio is to show whether certain sector is worth to be invested because P/E ratio can reflect the investment risk in this sector. The figure below shows distribution of P/E ratios in different sectors.
+\begin{center}
+    \includegraphics[scale=0.4]{PE_Ratio.png}
+\end{center}
+By comparing cross-sectional data, it is obvious that P/E of Real Estate Sector and Consumer Discretionary Sector are higher than others. The reason is that earning growth in the future is expected to grow fast in the two sectors or these sectors have some special advantages that guarantee long-term profitability with low risk. On the other hand, Financials Sector’s ratio is relatively low compared with other sectors, which may result from its high volatility so investors are reluctant to pay for it.\\
+We also notice that the ratio in Energy Sector surged in 2015, which is related to some changes in the sector. The end of the oil age and emergence of alternative energy have reduced the earnings of the original sector. As a result, its relative price becomes higher than before. The change in EV/EBIT and EV/EBITDA is also due to this reason.
+
+\subsection{PB Ratio}
+The P/B ratio provides a valuable reality check for investors who are seeking growth at a reasonable price. For those sectors with more assets, their book value and market value are close, so P/B ratio is more useful when we analyze Real Estate sector and Financials sector. The figure below shows distribution of P/B ratio in different sectors.
+\begin{center}
+    \includegraphics[scale=0.4]{PB_Ratio.png}
+\end{center}
+As the picture shows, Consumer Discretionary sector and IT sector have higher P/B ratio while Financials sector and Energy sector have relatively low ratios. What’s more, Real Estate sector with high P/E ratio has relatively lower P/B ratio.
+
+\subsection{EV/Sales}
+EV/Sales can help investors better understand cost relative to unit sales and whether the company is overvalued or undervalued. If EV/Sales is relatively high, the company or sector is less attractive to investors. The figure below shows distribution of EV/Sales in different sectors.
+\begin{center}
+    \includegraphics[scale=0.4]{EV2Sales.png}
+\end{center}
+The result shows that Real Estate sector’s ratio is higher than others’, which means that index in this sector is overvalued. On the other hand, ratio in Energy sector is low, which can attract more investors.
+
+\subsection{EV/EBIT $\&$ EV/EBITDA}
+EV/EBIT and EV/EBITDA are independent of the capital structure of the company, whereas multiples like P/E ratio are impacted by financing decisions. Because of this reason, the two are the most commonly relied-upon multiples in relative valuation. However, one obvious distinction is that EV/EBIT considers depreciation and amortization. In some capital-intensive industries which have significant differences in D$\&$A, EV/EBIT may make it a more accurate measure of value. But in our analysis, there is no such difference in the comparison of these two ratios under different sectors.
+\begin{center}
+    \includegraphics[scale=0.4]{EV2EBIT.png}
+    \includegraphics[scale=0.4]{EV2EBITDA.png}
+\end{center}
+The results can show that Real Estate sector has a higher ratio. The three ratio EV/Sales, EV/EBIT and EV/EBITDA can give a consensus conclusion that Real Estate sector is overvalued in the market.
+
+\subsection{Dividend Yield}
+Dividend Yield is used to measure the amount of cash flow investors are getting back for each dollar. It is essentially the return on investment for a stock without any capital gains. The figure below shows distribution of Dividend yield in different sectors.
+\begin{center}
+    \includegraphics[scale=0.4]{Div_Yield.png}
+\end{center}
+The ratio in Communication Services sector is higher before 2018 while Energy sector’s ratio is higher after that time. This is because communication services sector took place a reorganization of S$\&$P500 index in 2018. It now includes at least eighteen companies from IT and Consumer Discretionary sectors. Due to this reshuffling, dividend yield of this sector is impacted.
+
+\subsection{Gross Margin}
+Gross margin equals net sales less the cost of goods sold (COGS). Net sales are equivalent to the total revenue from sales, and COGS is the direct cost associated with producing goods. By calculating gross margin, we could measure one company's retain revenue after subtracting the production cost. The higher the gross margin, the more capital a company retains, which it can then use to pay other costs or satisfy debt obligations. Generally, companies with good gross margins would have a relatively sustainable competitive advantage. By analyzing gross margin data across sectors, we may observe some sectors that have more stable development in the long run.
+\begin{center}
+    \includegraphics[scale=0.4]{gross_margin.png}
+\end{center}
+For our 11 sectors’ gross margin data, the line chart above shows significant differences between the sectors. Overall, each industry index is relatively flat on its own, and have gaps between each others. Utilities, Communication Services and Information Technology(IT) have been among the top spears for last 10 years, occupying the first, second and third positions respectively, all above 40$\%$. On the contrary, the energy sector has been an under-performer for the past decade, ranking at the bottom, with gross margins consistently below 20$\%$. Gross margins in the rest industries are concentrated in the 25$\%$-35$\%$ range and have not fluctuate much.\\
+At the same time, by observing the comparison of fluctuations between industries, it is not difficult to see that the gross margin fluctuations of the energy industry and the utilities industry maybe relatively high in the past decade, and their peaks correspond to each other. During 2016, the utilities industry grew significantly, while energy declined comparatively. The trend was even more pronounced in 2020, with utilities reaching its highest level and the energy industry fell to the bottom.\\
+Generally, the gross margin feature maybe a significant indicator for Utilities, Communication Services and IT sectors. And our conjecture about the correlation between utilities and energy sectors will need further observation and verification.
+
+\subsection{Operating Margin $\&$ Profit Margin:}
+Operating margin equals operating income divided by revenue, it is a profitability ratio measuring revenue after covering operating and non-operating expenses of a business. And profit margin measures the profit ratio after paying for variable costs of production. It is calculated by the formula: 
+$$\text{Profit Margin} = \frac{(\text{Total Revenue - Total Expenses )}}{ \text{Total Revenue}}$$
+Both operating margin and profit margin are used to gauge the degree of the company's activity makes money. Higher ratios are generally better, illustrating the company is efficient in its operations and is good at turning sales into profits. In our analysis, there is not a very big difference in the comparison of these two ratios under different sectors, which is determined by their definition.
+\begin{center}
+    \includegraphics[scale=0.4]{operating_margin.png}
+    \includegraphics[scale=0.4]{profit_margin.png}
+\end{center}
+For these two ratios, Real Estate sector, IT sector and Financial sector have the top three high ratios.And Energy sector has the relatively lowest ratio. Also, both operating margin and profit margin for almost all sectors have similar trends in the last decade curves. This is attributed to the definition difference between the two features, and that's why the operating margin was slightly higher than the profit margin.\\
+Another thing that is worth to mentioning is that for Energy Sector, not just operating margin and profit margin, but also the gross margin, it always has the relatively lowest ratios and similar curve fluctuation, with sharp declines in 2016 and 2020. The two time nodes may consistent with some big revolution in the energy industry, which we will analyze later.
+
+
+\subsection{Return on Asset $\&$ Return on Equity}
+Return on equity (ROE) and return on assets (ROA) are two of the most important measures for evaluating how effectively a company’s management team is doing its job of managing the capital entrusted to it. ROE equals to generally net income divided by equity, while Return on Assets (ROA) is net income divided by average assets. So the primary differentiator between ROE and ROA is financial leverage or debt. ROE measures profitability and ROA is an efficiency measure of how well a company is using its assets.
+Investors may prefer to observe ROE, since equity represents the owner's interest in the business. Compared to other sources of fund, equity capital tends to be the most expensive source of funding and carries the largest risk premium of all financing options. Therefore, in our analysis, ROE may be a better feature that it could reflect the trend of market investment.
+\begin{center}
+    \includegraphics[scale=0.4]{roa.png}
+\end{center}
+As shown in the picture, IT sector has the highest ROA, the Consumer Staples sector and Consumer Discretionary sector also have a relatively higher ratio. In contrast, Financial sector has a lower ROA. 
+The past ten years, or even twenty years, has been an era of rapid development of information technology. And compared with traditional industry and commerce, information technology is more flexible in the time and form of investment assets, that's the reason why IT will have the highest ratio. Also for the the Consumer Staples sector and Consumer Discretionary sector，they are all industries with fast innovation and short production cycle. Generally, these three will have constantly higher ratio for the long run. Therefore, for these three industries, if the ROA indicator fluctuates significantly, it may have an impact on the investment trend.
+\begin{center}
+    \includegraphics[scale=0.40]{roe.png}
+\end{center}
+For ROE ratio, similarly, IT, Consumer Staples stay high, and Consumer Discretionary sectors is also at a slightly higher level, except that the IT sector lost its prominence in ROA ratio. 
+By comparing cross-sectional data, the Consumer Discretionary Sector and Industrials Sector have similar patterns in the last decade for both ROA and ROE ratios. They both have a low peak in 2020. It is conceivable that this is affected by the general environment of the epidemic.
+And as we mentioned before, the ROE and ROA curves of the energy sector still have a similar pattern, falling sharply in 2016 and 2020. In 2016, it was affected by changes in energy policy since 2015, reducing oil production while encouraging the development of clean and new energy. For 2020, we attribute this decline to the outbreak of the COVID-19 pandemic.
+\section{Prediction by Fundamental Factors}
+\subsection{Factor and Future Return}
+Having these fundamental data, next step is to find out what quantitative relationships they have to futures sector returns. For fundamental factors, they are usually exposed in the company report with annual, semi-annual, or quarterly frequency. Our fundamental factors for each sector are reported quarterly, leading to a problem that the sample size for each individual sector is very small. To have a better performance of the prediction model, we need to combine all the sectors together and make a uniformed and comparable large sample. We neutralized each factor cross-sectional for the factor to have a mean of 0 and standard deviation of 1. If $X_{i,t}$ denotes one specific factor exposure for $i$-th sector at time $t$, in this case would be at $t$-th quarter, then for each individual $t$ we have the neutralized exposure to be:
+$$X_{i,t}\text{ -Neutral}=\frac{X_{i,t}-\text{Mean}(X_{i,t}\text{  ,}i \text{ from } 1 \text{ to } 11)}{\text{Standard Deviation}(X_{i,t}\text{  ,}i \text{ from } 1 \text{ to } 11)}$$
+Then we used the next quarter's cross-sectional normalized return as the corresponding return. First, we want to have a general view of the relations. The scatter plots between neutralized factors and future returns are as following:
+\begin{center}
+\includegraphics[scale=0.5]{rPE.png}
+\includegraphics[scale=0.5]{rPB.png}\\
+\includegraphics[scale=0.5]{rOM.png}
+\includegraphics[scale=0.5]{rEV2EBIT.png}\\
+\includegraphics[scale=0.5]{rEV2EBITDA.png}
+\includegraphics[scale=0.5]{rEV2Sales.png}\\
+\includegraphics[scale=0.5]{rOM.png}
+\includegraphics[scale=0.5]{rDIV_Y.png}\\
+\includegraphics[scale=0.5]{rROE.png}
+\includegraphics[scale=0.5]{rROA.png}\\
+\end{center}
+From the scatter plots, the relations between all factors and their future returns cannot be well interpreted by simple linear models. However, it is very common in the financial field that the sample will have a very low signal-noise ratio.
+\subsection{Model Construction and Training}
+As we observed before, the relations between each factor and its future return cannot be interpreted very well by linear models. Also, we have no idea what model would exactly best fit the data. Therefore, converting prediction of future returns to a classification problem and fitting the training sample with a neural network model which has comparably good performance with non-linear relations would be a great start point. 
+\begin{center}
+    \begin{neuralnetwork}[height=10]
+            \newcommand{\x}[2]{$x_#2$}
+            \newcommand{\y}[2]{$\hat{y}_#2$}
+            \newcommand{\hfirst}[2]{\small $h^{(1)}_#2$}
+            \newcommand{\hsecond}[2]{\small $h^{(2)}_#2$}
+            \inputlayer[count=10, bias=false, title=Input\\layer, text=\x]
+            \hiddenlayer[count=5, bias=false, title=Hidden\\layer 1, text=\hfirst] \linklayers
+            \hiddenlayer[count=5, bias=false, title=Hidden\\layer 2, text=\hsecond] \linklayers
+            \outputlayer[count=2, title=Output\\layer, text=\y] \linklayers
+    \end{neuralnetwork}
+\end{center}
+Neural network takes a vector as the input, and goes to each of the neuron in the first hidden layer and gains new activation vectors which act as the input for next hidden layer. After the last hidden layer, neural network model would pass out the probability for each of the prediction class and we choose the one with the highest probability as the prediction. This process is called front propagation. After comparing the prediction to the actual results, we adjust the weights of the nodes by using back propagation for each training pair in the training samples. Also, we use the rectified linear unit function as the activation function for hidden layers and sigmoid function as the activation function for final output. Since we only have a sample of size 200, choosing quasi-Newton methods as the solver has better performance for small sample training.
+Then we need to construct the training, validation, and test sets. Since the fundamental factors are already neutralized (normalized) within each sector, we divide the sample data to 60$\%$, 20$\%$, 20$\%$ by convention. Without shuffling, we will have the historical data divided where test set contains the most recent data. For the corresponding output value, we assign 1 to samples with positive future return and 0 with negative returns.
+The complexity of neural network directly related to the number and sizes of hidden layers. For the purpose of avoiding overfitting or under-fitting, we need to find proper hyper parameters for neural network model. We start from a simple model with two layers. Let $N$ denote the number of nodes in each hidden layer, alpha is the hyper parameter for L2 regularization penalty function. With larger $N$, the model is more complex. If alpha increases, the penalty for large weights increases, which makes the model tend to be more simple. Considering our sample size is small, intuitively we need to focus more on the overfitting problem. For a range of alpha and $N$, we train the model using the training set data, and get the score for prediction on validation set. The score represents the probability of making a right prediction. Here are the results:
+    \begin{center}
+    \begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|}
+   \hline
+        N$\backslash$alpha & 3 & 2.5 & 2 & 1.5 & 1 & 0.5 & 0.25 & 0.1 & 0.01 & 0.001  \\ \hline
+        5 & 0.49 & 0.45 & 0.49 & 0.56 & 0.62 & 0.58 & 0.62 & 0.58 & 0.56 & 0.56  \\ \hline
+        6 & 0.51 & 0.55 & 0.45 & 0.49 & 0.51 & 0.49 & 0.47 & 0.56 & 0.56 & 0.56  \\ \hline
+        7 & 0.55 & 0.53 & 0.56 & 0.51 & 0.47 & 0.51 & 0.58 & 0.49 & 0.53 & 0.57  \\ \hline
+        8 & 0.51 & 0.47 & 0.55 & 0.51 & 0.51 & 0.64 & 0.6 & 0.6 & 0.59 & 0.6  \\ \hline
+        9 & 0.53 & 0.49 & 0.47 & 0.53 & 0.56 & 0.56 & 0.55 & 0.6 & 0.53 & 0.6  \\ \hline
+        10 & 0.55 & 0.53 & 0.53 & 0.53 & 0.51 & 0.51 & 0.53 & 0.6 & 0.6 & 0.56  \\ \hline
+        11 & 0.53 & 0.58 & 0.51 & 0.51 & 0.47 & 0.65 & 0.58 & 0.55 & 0.56 & 0.51  \\ \hline
+        12 & 0.55 & 0.55 & 0.56 & 0.51 & 0.49 & 0.51 & 0.54 & 0.53 & 0.53 & 0.53  \\ \hline
+        13 & 0.53 & 0.47 & 0.55 & 0.49 & 0.56 & 0.6 & 0.55 & 0.58 & 0.6 & 0.56  \\ \hline
+        14 & 0.49 & 0.51 & 0.56 & 0.53 & 0.55 & 0.53 & 0.49 & 0.51 & 0.6 & 0.53  \\ \hline
+        15 & 0.47 & 0.51 & 0.51 & 0.51 & 0.45 & 0.49 & 0.56 & 0.6 & 0.56 & 0.55 \\ \hline
+    \end{tabular}
+\end{center}
+To better understand how the hyperparameters influence model performance, we visualize the data by using $N$ and alpha as the bottom coordinates, and use the corresponding probability as the height.
+\begin{center}
+\includegraphics[scale=0.4]{hyperparameter_performance.png}
+\end{center}
+From the figure, we can tell that the model have several local optimal pairs. And the scores at the optimal points with relative large N values are also combined with small alpha values. For example, the combination of 14 nodes and alpha equals 0.01 has a local optimal score of 0.6. Since we are training with a small sample, using such a complex model with a high score is highly likely overfitting. Therefore, we start from the simple model by looking at models with 5 nodes model and check how the score varies with alpha. For model with 5 nodes, we see there are two local peaks with alpha equal to 1 and 0.25, then we pick the middle value 0.5 as the value of alpha considering the trade-off between variance and bias. 
+
+\subsection{Model Performance}
+Constructed and trained the model, next we would test the model by feeding a new data set to the model. On the test set, the score of the model is 0.64, which means the model predicts 64$\%$ of the results correctly. More detailed results are showed in the following table:
+
+\begin{table}[!ht]
+    \centering
+    \begin{tabular}{|l|c|c|}
+    \hline
+        ~ & Positive Return& Negative Return\\ \hline
+        Actual Sample & 27 & 28 \\ \hline
+        Prediction & 37 & 18 \\ \hline
+        Correct Prediction & 22 & 13 \\ \hline
+        Winning Rate & 0.59 & 0.72 \\ \hline
+    \end{tabular}
+\end{table}
+On the test set, we have a 0.59 winning rates on the positive predictions and 0.72 on the negative predictions, which gives an overall winning rate of 0.64.
+
+\subsection{Trade with Model Prediction}
+By using the predictions from the trained model, we used the data from validation set to get trade signals. Instead of having signals of 1 or 0 as the model's output, we choose the probability of the prediction output being 1, which is given by the activation sigmoid function. Then we will have a time series of the probability for each sector, and rank the probability from highest to lowest where the highest probability will have a rank 1. For each cross-sectional ranking, we equally-weighted long sectors with rank 1 to 3 and short sectors with rank 9 to 11 to construct a dollar-neutral portfolio. On the test set, which is from September in 2020 to September in 2021, we have a Sharpe ratio of 2.21. The cumulative return plot is following:
+\begin{center}
+\includegraphics[scale=0.4]{trade_on_model.png}
+\end{center}
+\subsection{Other Considerations}
+There are still issues that need to be considered carefully in the future. First is the factor neutralization. In previous model, we neutralized the factor exposure cross-sectionally, where the exposures reflect the relative level of factor exposure for one sector compared to other sectors at a given time. However, different sectors may have inner trends of higher exposures than others for some factors, especially for fundamental factor. What's more, we only have quarterly fundamental data available from 2017 and it is hard to implement time series normalization for each sector. Therefore, how to modify the factor exposures to make them comparable is a difficult problem. Secondly, as the sample size is small, the model might not be applicable on a wider range of time since we only trained and tested on the most recent five years. One possible way to improve this model is to use daily factors such as volume, close price as input, and convert fundamental factors to daily frequency by the corresponding quarter. Then we would have a sample size of approximately 1250 for each sector and over 13000 samples for training. However, the model might depends more on the daily factors rather than fundamental factors since their exposures would be the same value for each quarter.
+\section{References}
+1. \textit{Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency} Narasimhan Jegadeesh; Sheridan Titman
+The Journal of Finance, Vol. 48, No. 1. (Mar., 1993), pp. 65-91.\\
+2.The Global Industry Classification Standard, MSCI (1999)
+\end{document}

arXiv_src_0001_001/0001/astro-ph0001001/flat_rot.eps

@@ -0,0 +1,1412 @@
+%!PS-Adobe-2.0 EPSF-2.0
+%%Creator: SM
+%%BoundingBox: 18 144 592 718
+%%DocumentFonts: Helvetica
+%%EndComments
+ 20 dict begin
+72 300 div dup scale
+1 setlinejoin 0 setlinecap
+/Helvetica findfont 55 scalefont setfont 
+/B {CS newpath moveto} def
+/F {moveto 0 setlinecap} def
+/C {CS M 1 1 3 {pop 3 1 roll 255 div} for SET_COLOUR} def
+/CS {currentpoint stroke} def
+/CF {currentpoint eofill} def
+/L {lineto} def /l {rlineto} def /M {moveto} def /m {rmoveto} def
+/P {moveto 0 1 rlineto stroke} def
+/T {currentlinecap exch 1 setlinecap show setlinecap} def
+errordict /nocurrentpoint {pop 0 0 M currentpoint} put
+/SET_COLOUR statusdict begin /processcolors where
+  {pop processcolors 1 gt} {false} ifelse
+  {(setrgbcolor)} {(pop pop pop)} ifelse cvx end def
+ 80 600 translate
+gsave
+CS [] 0 setdash M
+CS M 1 dup setlinewidth
+/P [ /moveto cvx 0 5 -1 roll .05 add dup -2 div 0 exch /rmoveto cvx /rlineto cvx /stroke cvx ] cvx def
+ 0 0 0 C
+CS [] 0 setdash M
+255 255 M 2006 0 l
+255 255 M 0 37 l
+-15 -121 m 240 171 F (0) T
+361 255 M 0 19 l
+105 -19 m 0 19 l
+106 -19 m 0 19 l
+106 -19 m 0 37 l
+-31 -121 m 647 171 F (20) T
+783 255 M 0 19 l
+106 -19 m 0 19 l
+105 -19 m 0 19 l
+106 -19 m 0 37 l
+-31 -121 m 1069 171 F (40) T
+1205 255 M 0 19 l
+106 -19 m 0 19 l
+106 -19 m 0 19 l
+105 -19 m 0 37 l
+-31 -121 m 1491 171 F (60) T
+1628 255 M 0 19 l
+105 -19 m 0 19 l
+106 -19 m 0 19 l
+105 -19 m 0 37 l
+-31 -121 m 1913 171 F (80) T
+2050 255 M 0 19 l
+106 -19 m 0 19 l
+105 -19 m 0 19 l
+255 2261 M 2006 0 l
+255 2261 M 0 -36 l
+106 36 m 0 -18 l
+105 18 m 0 -18 l
+106 18 m 0 -18 l
+106 18 m 0 -36 l
+105 36 m 0 -18 l
+106 18 m 0 -18 l
+105 18 m 0 -18 l
+106 18 m 0 -36 l
+105 36 m 0 -18 l
+106 18 m 0 -18 l
+106 18 m 0 -18 l
+105 18 m 0 -36 l
+106 36 m 0 -18 l
+105 18 m 0 -18 l
+106 18 m 0 -18 l
+105 18 m 0 -36 l
+106 36 m 0 -18 l
+106 18 m 0 -18 l
+105 18 m 0 -18 l
+255 255 M 0 2006 l
+255 255 M 34 0 l
+118 237 M 9 28 m 3 1 l
+5 5 l
+0 -34 l
+-1 33 m 0 -33 l
+-7 0 m 15 0 l
+22 34 m -5 -1 l
+-3 -5 l
+-1 -8 l
+0 -5 l
+1 -8 l
+3 -5 l
+5 -2 l
+3 0 l
+5 2 l
+3 5 l
+2 8 l
+0 5 l
+-2 8 l
+-3 5 l
+-5 1 l
+-3 0 l
+-3 -1 l
+-2 -2 l
+-1 -3 l
+-2 -8 l
+0 -5 l
+2 -8 l
+1 -3 l
+2 -2 l
+3 -2 l
+3 0 m 4 2 l
+1 2 l
+2 3 l
+1 8 l
+0 5 l
+-1 8 l
+-2 3 l
+-1 2 l
+-4 1 l
+19 -9 m 17 0 l
+8 7 m 1 -1 l
+-1 -1 l
+-1 1 l
+0 1 l
+1 2 l
+1 1 l
+3 1 l
+3 0 l
+3 -1 l
+1 -2 l
+0 -3 l
+-1 -2 l
+-3 -1 l
+-3 0 l
+3 9 m 2 -1 l
+1 -2 l
+0 -3 l
+-1 -2 l
+-2 -1 l
+2 -1 l
+2 -1 l
+1 -2 l
+0 -3 l
+-1 -2 l
+-1 -1 l
+-3 -1 l
+-3 0 l
+-3 1 l
+-1 1 l
+-1 2 l
+0 1 l
+1 1 l
+1 -1 l
+-1 -1 l
+10 6 m 1 -3 l
+0 -3 l
+-1 -2 l
+-1 -1 l
+-2 -1 l
+37 123 m 17 0 l
+-17 71 m 17 0 l
+-17 50 m 17 0 l
+-17 39 m 17 0 l
+-17 31 m 17 0 l
+-17 27 m 17 0 l
+-17 24 m 17 0 l
+-17 20 m 17 0 l
+-17 18 m 34 0 l
+118 639 M 9 27 m 3 2 l
+5 4 l
+0 -33 l
+-1 32 m 0 -32 l
+-7 0 m 15 0 l
+22 33 m -5 -1 l
+-3 -5 l
+-1 -8 l
+0 -5 l
+1 -8 l
+3 -5 l
+5 -1 l
+3 0 l
+5 1 l
+3 5 l
+2 8 l
+0 5 l
+-2 8 l
+-3 5 l
+-5 1 l
+-3 0 l
+-3 -1 l
+-2 -2 l
+-1 -3 l
+-2 -8 l
+0 -5 l
+2 -8 l
+1 -3 l
+2 -2 l
+3 -1 l
+3 0 m 4 1 l
+1 2 l
+2 3 l
+1 8 l
+0 5 l
+-1 8 l
+-2 3 l
+-1 2 l
+-4 1 l
+19 -9 m 17 0 l
+8 7 m 1 -1 l
+-1 -1 l
+-1 1 l
+0 1 l
+1 2 l
+1 1 l
+3 1 l
+3 0 l
+3 -1 l
+1 -1 l
+1 -2 l
+0 -2 l
+-1 -2 l
+-3 -1 l
+-4 -2 l
+-2 -1 l
+-2 -2 l
+-1 -3 l
+0 -3 l
+8 20 m 2 -1 l
+1 -1 l
+1 -2 l
+0 -2 l
+-1 -2 l
+-3 -1 l
+-3 -2 l
+-5 -7 m 1 1 l
+2 0 l
+4 -2 l
+3 0 l
+2 1 l
+1 1 l
+-10 0 m 4 -3 l
+4 0 l
+1 1 l
+1 2 l
+0 2 l
+32 118 m 17 0 l
+-17 71 m 17 0 l
+-17 50 m 17 0 l
+-17 39 m 17 0 l
+-17 32 m 17 0 l
+-17 26 m 17 0 l
+-17 24 m 17 0 l
+-17 20 m 17 0 l
+-17 19 m 34 0 l
+133 1030 M 133 1030 F (0.1) T
+255 1178 M 17 0 l
+-17 71 m 17 0 l
+-17 50 m 17 0 l
+-17 39 m 17 0 l
+-17 32 m 17 0 l
+-17 27 m 17 0 l
+-17 23 m 17 0 l
+-17 20 m 17 0 l
+-17 19 m 34 0 l
+-94 -28 m 195 1431 F (1) T
+255 1580 M 17 0 l
+-17 70 m 17 0 l
+-17 50 m 17 0 l
+-17 39 m 17 0 l
+-17 32 m 17 0 l
+-17 27 m 17 0 l
+-17 23 m 17 0 l
+-17 21 m 17 0 l
+-17 18 m 34 0 l
+143 1842 M 9 27 m 4 2 l
+4 5 l
+0 -34 l
+-1 32 m 0 -32 l
+-7 0 m 15 0 l
+22 34 m -5 -2 l
+-3 -5 l
+-1 -8 l
+0 -4 l
+1 -8 l
+3 -5 l
+5 -2 l
+3 0 l
+5 2 l
+3 5 l
+2 8 l
+0 4 l
+-2 8 l
+-3 5 l
+-5 2 l
+-3 0 l
+-3 -2 l
+-2 -1 l
+-1 -4 l
+-2 -8 l
+0 -4 l
+2 -8 l
+1 -4 l
+2 -1 l
+3 -2 l
+3 0 m 4 2 l
+1 1 l
+2 4 l
+1 8 l
+0 4 l
+-1 8 l
+-2 4 l
+-1 1 l
+-4 2 l
+21 -2 m 2 1 l
+2 3 l
+0 -21 l
+-1 20 m 0 -20 l
+-3 0 m 8 0 l
+34 124 m 17 0 l
+-17 70 m 17 0 l
+-17 50 m 17 0 l
+-17 39 m 17 0 l
+-17 32 m 17 0 l
+-17 27 m 17 0 l
+-17 23 m 17 0 l
+-17 21 m 17 0 l
+-17 18 m 34 0 l
+143 2243 M 9 28 m 4 1 l
+4 5 l
+0 -34 l
+-1 32 m 0 -32 l
+-7 0 m 15 0 l
+22 34 m -5 -2 l
+-3 -4 l
+-1 -8 l
+0 -5 l
+1 -8 l
+3 -5 l
+CS M
+5 -2 l
+3 0 l
+5 2 l
+3 5 l
+2 8 l
+0 5 l
+-2 8 l
+-3 4 l
+-5 2 l
+-3 0 l
+-3 -2 l
+-2 -1 l
+-1 -3 l
+-2 -8 l
+0 -5 l
+2 -8 l
+1 -4 l
+2 -1 l
+3 -2 l
+3 0 m 4 2 l
+1 1 l
+2 4 l
+1 8 l
+0 5 l
+-1 8 l
+-2 3 l
+-1 1 l
+-4 2 l
+19 -2 m 1 -1 l
+-1 -1 l
+-1 1 l
+0 1 l
+1 2 l
+1 1 l
+3 1 l
+3 0 l
+3 -1 l
+1 -1 l
+1 -2 l
+0 -2 l
+-1 -2 l
+-3 -2 l
+-4 -2 l
+-2 -1 l
+-2 -2 l
+-1 -3 l
+0 -2 l
+8 20 m 2 -1 l
+1 -1 l
+1 -2 l
+0 -2 l
+-1 -2 l
+-3 -2 l
+-3 -2 l
+-5 -6 m 1 0 l
+2 0 l
+4 -1 l
+3 0 l
+2 1 l
+1 0 l
+-10 0 m 4 -2 l
+4 0 l
+1 1 l
+1 1 l
+0 2 l
+2261 255 M 0 2006 l
+2261 255 M -34 0 l
+34 121 m -17 0 l
+17 71 m -17 0 l
+17 50 m -17 0 l
+17 39 m -17 0 l
+17 31 m -17 0 l
+17 27 m -17 0 l
+17 24 m -17 0 l
+17 20 m -17 0 l
+17 18 m -34 0 l
+34 121 m -17 0 l
+17 71 m -17 0 l
+17 50 m -17 0 l
+17 39 m -17 0 l
+17 32 m -17 0 l
+17 26 m -17 0 l
+17 24 m -17 0 l
+17 20 m -17 0 l
+17 19 m -34 0 l
+34 120 m -17 0 l
+17 71 m -17 0 l
+17 50 m -17 0 l
+17 39 m -17 0 l
+17 32 m -17 0 l
+17 27 m -17 0 l
+17 23 m -17 0 l
+17 20 m -17 0 l
+17 19 m -34 0 l
+34 121 m -17 0 l
+17 70 m -17 0 l
+17 50 m -17 0 l
+17 39 m -17 0 l
+17 32 m -17 0 l
+17 27 m -17 0 l
+17 23 m -17 0 l
+17 21 m -17 0 l
+17 18 m -34 0 l
+34 121 m -17 0 l
+17 70 m -17 0 l
+17 50 m -17 0 l
+17 39 m -17 0 l
+17 32 m -17 0 l
+17 27 m -17 0 l
+17 23 m -17 0 l
+17 21 m -17 0 l
+17 18 m -34 0 l
+CS [] 0 setdash M
+406 255 M 22 36 l
+105 86 l
+104 61 l
+105 47 l
+104 40 l
+105 34 l
+104 31 l
+105 28 l
+104 26 l
+104 25 l
+105 24 l
+104 24 l
+105 23 l
+104 22 l
+105 23 l
+104 24 l
+105 24 l
+104 24 l
+324 495 M 104 165 l
+105 87 l
+104 60 l
+105 48 l
+104 39 l
+105 34 l
+104 31 l
+105 28 l
+104 27 l
+104 25 l
+105 24 l
+104 23 l
+105 23 l
+104 23 l
+105 23 l
+104 23 l
+105 24 l
+104 25 l
+0 -44 m -35 35 l
+35 -86 m -77 77 l
+77 -129 m -119 119 l
+119 -170 m -161 160 l
+161 -212 m -203 203 l
+203 -255 m -245 245 l
+245 -296 m -287 287 l
+259 -310 m -301 300 l
+259 -310 m -301 301 l
+259 -311 m -302 302 l
+260 -312 m -302 303 l
+260 -312 m -302 302 l
+260 -312 m -303 303 l
+261 -312 m -303 303 l
+261 -313 m -303 303 l
+261 -312 m -303 303 l
+260 -312 m -302 303 l
+260 -312 m -302 302 l
+260 -312 m -302 302 l
+259 -311 m -301 302 l
+259 -311 m -301 301 l
+259 -310 m -301 300 l
+259 -310 m -300 300 l
+257 -309 m -298 298 l
+256 -308 m -297 298 l
+255 -307 m -296 296 l
+254 -306 m -295 295 l
+254 -305 m -294 294 l
+252 -304 m -292 292 l
+251 -302 m -291 290 l
+250 -300 m -289 288 l
+247 -299 m -286 286 l
+246 -297 m -284 284 l
+243 -295 m -281 281 l
+241 -292 m -278 278 l
+238 -289 m -275 274 l
+235 -286 m -270 270 l
+231 -282 m -267 266 l
+228 -279 m -262 262 l
+223 -275 m -256 256 l
+218 -270 m -251 251 l
+214 -265 m -246 246 l
+209 -261 m -237 237 l
+201 -253 m -229 230 l
+194 -246 m -222 222 l
+187 -238 m -212 212 l
+179 -231 m -199 199 l
+167 -218 m -187 186 l
+154 -205 m -174 174 l
+146 -197 m -166 165 l
+137 -189 m -153 154 l
+125 -177 m -125 125 l
+99 -151 m -99 100 l
+75 -127 m -75 75 l
+23 -75 m -23 23 l
+CS M 2 100 mul 72 div dup setlinewidth
+/P [ /moveto cvx 0 5 -1 roll .05 add dup -2 div 0 exch /rmoveto cvx /rlineto cvx /stroke cvx ] cvx def
+0 128 m 104 165 l
+105 87 l
+104 60 l
+105 48 l
+104 39 l
+105 35 l
+104 30 l
+105 28 l
+104 27 l
+104 25 l
+105 24 l
+104 23 l
+105 23 l
+104 23 l
+105 23 l
+104 23 l
+105 24 l
+104 25 l
+CS M 3 100 mul 72 div dup setlinewidth
+/P [ /moveto cvx 0 5 -1 roll .05 add dup -2 div 0 exch /rmoveto cvx /rlineto cvx /stroke cvx ] cvx def
+255 1680 M 30 4 l
+30 3 l
+30 4 l
+30 4 l
+30 3 l
+29 4 l
+30 4 l
+30 4 l
+30 5 l
+30 4 l
+30 4 l
+30 5 l
+29 4 l
+30 5 l
+30 5 l
+30 5 l
+30 5 l
+30 5 l
+30 5 l
+29 5 l
+30 6 l
+30 5 l
+30 6 l
+30 6 l
+30 6 l
+30 5 l
+29 7 l
+30 6 l
+30 6 l
+30 7 l
+30 6 l
+30 7 l
+30 7 l
+29 7 l
+30 7 l
+30 7 l
+30 8 l
+30 7 l
+30 8 l
+30 8 l
+29 8 l
+30 8 l
+30 8 l
+30 9 l
+30 8 l
+30 9 l
+30 10 l
+29 9 l
+30 9 l
+30 10 l
+30 10 l
+30 10 l
+30 11 l
+30 10 l
+29 11 l
+30 12 l
+30 11 l
+30 12 l
+30 12 l
+30 12 l
+30 13 l
+29 13 l
+30 14 l
+30 14 l
+30 14 l
+30 15 l
+30 15 l
+5 3 l
+CS [32 24] 0 setdash M
+255 1276 M 30 3 l
+30 2 l
+30 2 l
+30 3 l
+30 2 l
+29 3 l
+30 3 l
+30 2 l
+30 3 l
+30 3 l
+30 3 l
+CS M
+30 3 l
+29 3 l
+30 3 l
+30 3 l
+30 4 l
+30 3 l
+30 3 l
+30 4 l
+29 3 l
+30 4 l
+30 4 l
+30 3 l
+30 4 l
+30 4 l
+30 4 l
+29 4 l
+30 4 l
+30 5 l
+30 4 l
+30 4 l
+30 5 l
+30 4 l
+29 5 l
+30 5 l
+30 4 l
+30 5 l
+30 5 l
+30 6 l
+30 5 l
+29 5 l
+30 6 l
+30 5 l
+30 6 l
+30 6 l
+30 6 l
+30 6 l
+29 6 l
+30 6 l
+30 7 l
+30 6 l
+30 7 l
+30 7 l
+30 7 l
+29 8 l
+30 7 l
+30 8 l
+30 8 l
+30 8 l
+30 8 l
+30 9 l
+29 8 l
+30 9 l
+30 10 l
+30 9 l
+30 10 l
+30 10 l
+5 2 l
+CS [] 0 setdash M
+CS M 1 dup setlinewidth
+/P [ /moveto cvx 0 5 -1 roll .05 add dup -2 div 0 exch /rmoveto cvx /rlineto cvx /stroke cvx ] cvx def
+255 575 M 30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+5 0 l
+255 1155 M 30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+29 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+30 0 l
+5 0 l
+CS [] 0 setdash M
+982 73 M 56 25 m 50 0 l
+19 33 m 20 -58 l
+-17 58 m 17 -50 l
+19 50 m -19 -58 l
+-25 58 m 16 0 l
+17 0 m 16 0 l
+78 11 m -5 -5 l
+-6 -9 l
+-6 -11 l
+-2 -14 l
+0 -11 l
+2 -14 l
+6 -11 l
+6 -8 l
+5 -6 l
+-5 84 m -6 -11 l
+-3 -9 l
+-3 -14 l
+0 -11 l
+3 -14 l
+3 -8 l
+6 -11 l
+27 72 m 0 -58 l
+3 58 m 0 -58 l
+28 39 m -28 -28 l
+14 11 m 16 -22 l
+-19 22 m 17 -22 l
+-39 58 m 11 0 l
+19 -19 m 17 0 l
+-47 -39 m 19 0 l
+11 0 m 17 0 l
+19 39 m 0 -39 l
+3 39 m 0 -39 l
+0 30 m 6 6 l
+8 3 l
+6 0 l
+8 -3 l
+3 -6 l
+0 -30 l
+-11 39 m 5 -3 l
+3 -6 l
+0 -30 l
+3 30 m 5 6 l
+9 3 l
+5 0 l
+8 -3 l
+3 -6 l
+0 -30 l
+-11 39 m 6 -3 l
+2 -6 l
+0 -30 l
+-69 39 m 11 0 l
+-11 -39 m 20 0 l
+11 0 m 19 0 l
+11 0 m 20 0 l
+60 69 m 1393 53 L
+91 53 m 3 6 l
+0 -11 l
+-3 5 l
+-3 3 l
+-5 3 l
+-11 0 l
+-6 -3 l
+-3 -3 l
+0 -5 l
+3 -3 l
+6 -3 l
+14 -5 l
+5 -3 l
+3 -3 l
+-31 19 m 3 -2 l
+6 -3 l
+14 -6 l
+5 -2 l
+3 -3 l
+0 -9 l
+-3 -2 l
+-5 -3 l
+-12 0 l
+-5 3 l
+-3 2 l
+-3 6 l
+0 -11 l
+3 5 l
+44 64 m 6 -5 l
+6 -9 l
+5 -11 l
+3 -14 l
+0 -11 l
+-3 -14 l
+-5 -11 l
+-6 -8 l
+-6 -6 l
+6 84 m 6 -11 l
+2 -9 l
+3 -14 l
+0 -11 l
+-3 -14 l
+-2 -8 l
+-6 -11 l
+CS [] 0 setdash M
+CS [] 0 setdash M
+77 948 M -34 80 m -5 3 l
+11 0 l
+-6 -3 l
+-3 -3 l
+-2 -5 l
+0 -11 l
+2 -6 l
+CS M
+3 -3 l
+6 0 l
+3 3 l
+2 6 l
+6 14 l
+3 5 l
+2 3 l
+-19 -31 m 3 3 l
+3 6 l
+5 14 l
+3 5 l
+3 3 l
+8 0 l
+3 -3 l
+3 -5 l
+0 -12 l
+-3 -5 l
+-3 -3 l
+-6 -3 l
+12 0 l
+-6 3 l
+-25 78 m 3 -3 l
+3 3 l
+-3 3 l
+-3 0 l
+-6 -6 l
+-2 -6 l
+0 -8 l
+2 -8 l
+6 -6 l
+8 -2 l
+6 0 l
+8 2 l
+6 6 l
+3 8 l
+0 6 l
+-3 8 l
+-6 6 l
+-30 -20 m 2 -5 l
+6 -6 l
+8 -3 l
+6 0 l
+8 3 l
+6 6 l
+3 5 l
+-34 42 m 3 0 l
+0 -3 l
+-3 0 l
+-3 3 l
+-2 5 l
+0 11 l
+2 6 l
+3 3 l
+6 2 l
+19 0 l
+6 3 l
+3 3 l
+-34 -8 m 25 0 l
+6 2 l
+3 6 l
+0 3 l
+-28 -11 m 3 -3 l
+2 -17 l
+3 -8 l
+6 -3 l
+5 0 l
+6 3 l
+3 8 l
+0 9 l
+-3 5 l
+-6 6 l
+-14 -20 m 3 -5 l
+6 -3 l
+5 0 l
+6 3 l
+3 5 l
+18 1161 M 59 0 l
+18 1164 M 59 0 l
+18 1153 M 0 11 l
+59 -11 m 0 19 l
+-23 17 m 0 33 l
+-5 0 l
+-6 -3 l
+-3 -2 l
+-2 -6 l
+0 -8 l
+2 -9 l
+6 -5 l
+8 -3 l
+6 0 l
+8 3 l
+6 5 l
+3 9 l
+0 5 l
+-3 9 l
+-6 5 l
+-14 -3 m -8 0 l
+-6 -2 l
+-2 -14 m 2 -6 l
+6 -5 l
+8 -3 l
+6 0 l
+8 3 l
+6 5 l
+3 6 l
+-34 63 m -5 3 l
+11 0 l
+-6 -3 l
+-3 -2 l
+-2 -6 l
+0 -11 l
+2 -5 l
+3 -3 l
+6 0 l
+3 3 l
+2 5 l
+6 14 l
+3 5 l
+2 3 l
+-19 -30 m 3 3 l
+3 5 l
+5 14 l
+3 5 l
+3 3 l
+8 0 l
+3 -3 l
+3 -5 l
+0 -11 l
+-3 -6 l
+-3 -2 l
+-6 -3 l
+12 0 l
+-6 3 l
+7 1352 M 6 -5 l
+8 -6 l
+11 -5 l
+14 -3 l
+11 0 l
+14 3 l
+11 5 l
+8 6 l
+6 5 l
+13 1347 M 11 -6 l
+8 -2 l
+14 -3 l
+11 0 l
+14 3 l
+8 2 l
+11 6 l
+18 1375 M 59 0 l
+18 1377 M 59 0 l
+38 1405 M 27 -28 l
+-11 14 m 23 17 l
+-23 -20 m 23 17 l
+18 1366 M 0 11 l
+20 20 m 0 16 l
+39 -47 m 0 20 l
+0 11 m 0 16 l
+38 1433 M 58 0 l
+38 1436 M 58 0 l
+46 1436 M -6 5 l
+-2 6 l
+0 5 l
+2 8 l
+6 6 l
+8 3 l
+6 0 l
+8 -3 l
+6 -6 l
+3 -8 l
+0 -5 l
+-3 -6 l
+-6 -5 l
+-30 16 m 2 6 l
+6 5 l
+8 3 l
+6 0 l
+8 -3 l
+6 -5 l
+3 -6 l
+38 1424 M 0 12 l
+58 -12 m 0 20 l
+46 1519 M 3 -3 l
+3 3 l
+-3 2 l
+-3 0 l
+-6 -5 l
+-2 -6 l
+0 -8 l
+2 -8 l
+6 -6 l
+8 -3 l
+6 0 l
+8 3 l
+6 6 l
+3 8 l
+0 6 l
+-3 8 l
+-6 5 l
+-30 -19 m 2 -5 l
+6 -6 l
+8 -3 l
+6 0 l
+8 3 l
+6 6 l
+3 5 l
+7 1538 M 6 6 l
+8 5 l
+11 6 l
+14 2 l
+11 0 l
+14 -2 l
+11 -6 l
+8 -5 l
+6 -6 l
+13 1544 M 11 5 l
+8 3 l
+14 3 l
+11 0 l
+14 -3 l
+8 -3 l
+11 -5 l
+CS [] 0 setdash M
+554 1779 M CS [] 0 setdash M
+645 1880 M 9 0 l
+10 -5 l
+5 -5 l
+5 -10 l
+28 -67 l
+5 -9 l
+5 -5 l
+-58 101 m 10 -10 l
+5 -10 l
+28 -67 l
+5 -9 l
+10 -5 l
+5 0 l
+-39 67 m -38 -67 l
+38 67 m -33 -67 l
+81 0 m CS [] 0 setdash M
+491 1362 M CS [] 0 setdash M
+668 1463 M 0 -101 l
+5 101 m 0 -101 l
+-20 101 m 34 0 l
+-34 -101 m 72 0 l
+0 29 l
+-5 -29 l
+10 0 m CS [] 0 setdash M
+491 907 M CS [] 0 setdash M
+668 1008 M 0 -101 l
+5 101 m 0 -101 l
+62 101 m -62 -63 l
+24 20 m 38 -58 l
+-43 58 m 38 -58 l
+-76 101 m 33 0 l
+29 0 m 29 0 l
+-91 -101 m 33 0 l
+29 0 m 29 0 l
+16 72 m 52 0 l
+28 23 m 6 3 l
+9 9 l
+0 -61 l
+-3 58 m 0 -58 l
+-12 0 m 26 0 l
+15 -46 m CS [] 0 setdash M
+1450 1195 M CS [] 0 setdash M
+1505 1253 M 11 -58 l
+-8 58 m 8 -44 l
+11 44 m -11 -58 l
+11 58 m 11 -58 l
+-8 58 m 8 -44 l
+11 44 m -11 -58 l
+-41 58 m 19 0 l
+25 0 m 17 0 l
+16 -25 m 0 -2 l
+-2 0 l
+0 2 l
+2 3 l
+6 3 l
+11 0 l
+5 -3 l
+3 -3 l
+3 -5 l
+0 -20 l
+3 -5 l
+3 -3 l
+-9 33 m 0 -25 l
+3 -5 l
+6 -3 l
+2 0 l
+-11 28 m -3 -3 l
+-16 -3 l
+-8 -3 l
+-3 -5 l
+0 -6 l
+3 -5 l
+8 -3 l
+8 0 l
+6 3 l
+5 5 l
+-19 14 m -6 -3 l
+-2 -5 l
+0 -6 l
+2 -5 l
+6 -3 l
+50 39 m 0 -39 l
+2 39 m 0 -39 l
+0 22 m 3 9 l
+CS M
+6 5 l
+5 3 l
+9 0 l
+2 -3 l
+0 -3 l
+-2 -2 l
+-3 2 l
+3 3 l
+-34 3 m 11 0 l
+-11 -39 m 20 0 l
+36 39 m 0 -39 l
+3 39 m 0 -39 l
+0 31 m 5 5 l
+8 3 l
+6 0 l
+8 -3 l
+3 -5 l
+0 -31 l
+-11 39 m 6 -3 l
+2 -5 l
+0 -31 l
+3 31 m 6 5 l
+8 3 l
+5 0 l
+9 -3 l
+3 -5 l
+0 -31 l
+-12 39 m 6 -3 l
+3 -5 l
+0 -31 l
+-69 39 m 11 0 l
+-11 -39 m 19 0 l
+11 0 m 19 0 l
+11 0 m 20 0 l
+64 58 m 0 -58 l
+2 58 m 0 -58 l
+34 58 m 0 -58 l
+2 58 m 0 -58 l
+-47 58 m 20 0 l
+16 0 m 20 0 l
+-45 -27 m 34 0 l
+-45 -31 m 20 0 l
+16 0 m 20 0 l
+19 58 m 0 -58 l
+3 58 m 0 -58 l
+-11 58 m 19 0 l
+-19 -58 m 19 0 l
+6 0 m CS [] 0 setdash M
+1450 615 M CS [] 0 setdash M
+1541 665 M 3 -8 l
+0 17 l
+-3 -9 l
+-6 6 l
+-8 3 l
+-5 0 l
+-9 -3 l
+-5 -6 l
+-3 -5 l
+-3 -9 l
+0 -13 l
+3 -9 l
+3 -5 l
+5 -6 l
+9 -3 l
+5 0 l
+8 3 l
+6 6 l
+3 5 l
+-22 45 m -6 -3 l
+-5 -6 l
+-3 -5 l
+-3 -9 l
+0 -13 l
+3 -9 l
+3 -5 l
+5 -6 l
+6 -3 l
+55 39 m -8 -3 l
+-6 -5 l
+-3 -8 l
+0 -6 l
+3 -8 l
+6 -6 l
+8 -3 l
+6 0 l
+8 3 l
+5 6 l
+3 8 l
+0 6 l
+-3 8 l
+-5 5 l
+-8 3 l
+-6 0 l
+-5 -3 l
+-6 -5 l
+-3 -8 l
+0 -6 l
+3 -8 l
+6 -6 l
+5 -3 l
+6 0 m 5 3 l
+6 6 l
+2 8 l
+0 6 l
+-2 8 l
+-6 5 l
+-5 3 l
+38 20 m 0 -59 l
+3 59 m 0 -59 l
+-11 59 m 11 0 l
+-11 -59 m 19 0 l
+48 59 m 0 -59 l
+2 59 m 0 -59 l
+-2 31 m -6 5 l
+-5 3 l
+-6 0 l
+-8 -3 l
+-6 -5 l
+-3 -8 l
+0 -6 l
+3 -8 l
+6 -6 l
+8 -3 l
+6 0 l
+5 3 l
+6 6 l
+-17 30 m -6 -3 l
+-5 -5 l
+-3 -8 l
+0 -6 l
+3 -8 l
+5 -6 l
+6 -3 l
+8 59 m 11 0 l
+-2 -59 m 11 0 l
+63 59 m 0 -59 l
+3 59 m 0 -59 l
+33 59 m 0 -59 l
+3 59 m 0 -59 l
+-47 59 m 20 0 l
+16 0 m 20 0 l
+-45 -28 m 33 0 l
+-44 -31 m 20 0 l
+16 0 m 20 0 l
+19 59 m 0 -59 l
+3 59 m 0 -59 l
+-11 59 m 19 0 l
+-19 -59 m 19 0 l
+6 0 m CS [] 0 setdash M
+stroke
+grestore
+showpage
+end

arXiv_src_0001_001/0001/astro-ph0001001/procl.sty

@@ -0,0 +1,95 @@
+\def\@aabuffer{}
+
+\def\author#1{\expandafter\def\expandafter\@aabuffer\expandafter
+{\@aabuffer\large#1\relax \par\vspace*{2ex}}}
+
+\def\address#1{\expandafter\def\expandafter\@aabuffer\expandafter
+{\@aabuffer\small#1\relax \par\vspace{.5ex}}}
+
+\def\maketitle{
+	\begin{center}
+	\vspace*{15ex}
+	{\Large\bf \@title \par}        
+	\vspace*{4ex}
+	\@aabuffer\relax
+	\end{center}
+\par
+\gdef\@aabuffer{}}
+
+\def\abstract#1{
+\begin{center}
+{\begin{minipage}{130truemm}
+                 \footnotesize\vspace*{3ex}
+                 \parindent0pt{\bf Abstract. }#1\par
+                 \end{minipage}}
+\end{center}
+                 \vspace*{0.5ex}\par}
+
+\fussy
+\flushbottom
+\frenchspacing
+\parindent15pt
+\oddsidemargin0mm         
+\evensidemargin0mm
+\topmargin15mm           
+\headheight3mm           
+\headsep4mm           
+\footskip10mm           
+\footheight4mm
+\textheight190.0truemm
+\textwidth130.0truemm
+
+\def\section{\@startsection{section}{1}{\z@}{-3.5ex plus -1ex minus 
+    -.2ex}{2.3ex plus .2ex}{\large\bf }}
+\def\subsection{\@startsection{subsection}{2}{\z@}{-3.25ex plus -1ex minus 
+   -.2ex}{1.5ex plus .2ex}{\bf }}
+
+\pagestyle{plain}
+
+\renewenvironment{thebibliography}[1]
+	{\begin{list}{\llap[\arabic{enumi}]}
+	{\usecounter{enumi}\setlength{\parsep}{0pt}
+	 \setlength{\itemsep}{0pt} \settowidth
+	{\labelwidth}{#1}\footnotesize\sloppy}}{\end{list}}
+
+\topsep0mm\parsep0mm\itemsep0mm
+
+\newcounter{arabiclistc}
+\newenvironment{arabiclist}
+	{\setcounter{arabiclistc}{0}
+	 \begin{list}{\arabic{arabiclistc}}
+	{\usecounter{arabiclistc}
+	 \setlength{\parsep}{0pt}
+	 \setlength{\itemsep}{0pt}}}{\end{list}}
+
+\setcounter{secnumdepth}{2}
+
+\def\baselinestretch{1.0}
+\ifx\selectfont\undefined
+\@normalsize\else\let\glb@currsize=\relax\selectfont
+\fi
+
+\ifx\selectfont\undefined
+\def\@singlespacing{%
+\def\baselinestretch{1}\ifx\@currsize\normalsize\@normalsize\else\@currsize\fi%
+}
+\else
+\def\@singlespacing{\def\baselinestretch{1}\let\glb@currsize=\relax\selectfont}
+\fi
+
+\def\figurename{Fig.}
+\def\tablename{Tab.}
+\long\def\@makecaption#1#2{
+   \vskip 10pt 
+   \setbox\@tempboxa\hbox{\footnotesize\bf #1.\rm~#2}
+   \ifdim \wd\@tempboxa >\hsize   
+       \leftskip 0pt plus 1fil 
+       \rightskip 0pt plus -1fil 
+       \parfillskip 0pt plus 2fil 
+       \footnotesize #1. #2\par   
+     \else                        
+       \hbox to\hsize{\hfil\box\@tempboxa\hfil}  
+   \fi}
+
+\def\references{\leftskip15pt\parindent-15pt\footnotesize}
+

arXiv_src_0001_001/0001/astro-ph0001001/procl.tex

@@ -0,0 +1,1505 @@
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+%			procl.tex			%
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+
+\documentstyle[procl]{article}
+
+\input{psfig.sty}
+
+% A useful Journal macro
+\def\Journal#1#2#3#4{(#1) {#2} {\bf #3}, #4}
+
+
+\newcommand{\be}{\begin{equation}}
+\newcommand{\ee}{\end{equation}}
+\newcommand{\ba}{\begin{eqnarray}}
+\newcommand{\ea}{\end{eqnarray}}
+\def \d {{\rm d}}
+\def \l {\langle}
+\def \r {\rangle}
+\def \ro {{\rm rck}}
+\def \de {{\rm des}}
+
+% Some useful journal names
+\def\AAp{\em Astron. Astrophys.}
+\def\AJ{\em Astron.~J.}
+\def\ApJ{\em Astrophys.~J.}
+\def\ApJL{\em Astrophys.~J., Lett.}
+\def\ApJSS{\em Astrophys.~J., Suppl. Ser.}
+\def\AaAp{\em Astron. Astrophys.}
+\def\ARFM{\em Annnu. Rev. Fluid Mech.}
+\def\ARAaAp{\em Annnu. Rev. Astron. Astrophys.}
+\def\MNRAS{\em Mon. Not. R.~Astron. Soc.}
+\def\Nat{\em Nature\/}
+\def\PASP{\em Publ. Astron. Soc. Pac.}
+\def\VA{\em Vistas Astron.}
+\def\IAUC{\em IAU Circ.}
+\def\SA{\em Sov. Astron.}
+\def\NC{\em Nuovo Cimento\/}
+\def\NPB{{\em Nucl. Phys.} B}
+\def\PLB{{\em Phys. Lett.}  B}
+\def\PRL{\em Phys. Rev. Lett.}
+\def\PRD{{\em Phys. Rev.} D}
+\def\ZPC{{\em Z. Phys.} C}
+
+
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+%			beginning of text			%
+%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
+
+
+\begin{document}
+
+
+\title{Statistics of Turbulence from Spectral-Line Data Cubes}
+
+\author{Alex Lazarian}
+
+\address{ Department of Astronomy, University of Wisconsin,
+Madison, USA; email: [email protected]}
+
+\maketitle
+
+\abstract{Emission in spectral lines can provide unique information
+on interstellar turbulence. Doppler shifts due to supersonic 
+motions contain information on turbulent velocity field which is
+otherwise difficult to measure. However, the problem of separation
+of velocity and density fluctuations is far from being trivial. Using atomic
+hydrogen (HI) as a test case, I review
+techniques applicable to emission line studies with the emphasis on those
+that can provide information on the underlying {\em power spectra} of
+velocity and density. I show 
+that recently developed mathematical machinery is promising
+for the purpose. Its application to HI
+ shows that in cold neutral hydrogen the velocity fluctuations dominate
+the small scale structures observed in spectral-line 
+data cubes and this result is
+very important for the interpretation of observational data, including
+the identification of clouds. Velocity fluctuations are shown to 
+dominate the formation of small scale structures that can be erroneously
+identified as diffuse clouds.
+One may argue that the HI
+data is consistent with the
+Goldreich-Shridhar picture of magnetohydrodynamic
+turbulence, but the cascade from the scales of several kpc that this
+interpretation involves
+does not fit well in the current paradigm of energy injection. 
+The
+issue whether magnetic field does make the turbulence anisotropic
+is still open, but if this is the case, I show that studies of emission
+lines can provide a reliable way of determining magnetic field direction.
+I discuss various techniques for studying 
+interstellar turbulence
+using emission lines, e.g. spectral correlation functions, genus
+statistics and principal component analysis.   
+}
+
+
+   \section{Introduction}
+
+
+The interstellar medium  is turbulent and the turbulence is crucial for
+understanding of various interstellar processes. Interstellar turbulence
+occurs in magnetized fluid and magnetic field establishes a connection
+between ISM phases (McKee \& Ostriker 1977) thus making the turbulent
+cascade much more complex and coupling together cosmic rays and gas. 
+Theoretical understanding of such a multiphase media with the
+injection of energy at different scales (Scalo 1987) is extremely
+challenging.
+
+In terms of the topic of the present meeting, turbulence is
+important both for accelerating cosmic rays and for their diffusion.
+Indeed, whatever mechanism of cosmic ray acceleration we consider,
+its understanding requires proper accounting for scattering
+of cosmic rays by turbulent magnetic field. The same is 
+true for the propagation of cosmic rays. For instance, if it were not 
+for magnetic field
+lines wandering, the diffusion of cosmic rays perpendicular to the
+magnetic
+field direction would be suppressed (see Jokipii 1999). Moreover,
+it is becoming clear that particle streaming along magnetic field lines
+is also substantially influenced by magnetic turbulence. 
+
+In view of a broader picture, turbulence is widely believed to
+be an important element of molecular cloud dynamics 
+and star formation process, although various 
+authors disagree on the degree of its importance (see discussion
+in Vazques-Semadeni
+\& Passot 1999). Undoubtedly turbulence is essential for heat 
+transfer in the interstellar medium. It has been recently
+suggested that turbulence
+is also 
+a key element to understanding various chemical reactions (Gredel 1999)
+and of the fundamental problem of MHD, namely, to the problem of
+fast magnetic reconnection (Lazarian \& Vishniac
+1999). This very limited and incomplete list of processes for which
+turbulence is essential explains the motivation behind the attempts
+to study interstellar turbulence.  
+
+
+Unfortunately interstellar turbulence remains a mystery in spite of all the
+attempts to study it. Substantial progress in numerical research (see
+Ostriker 1999, Vazquez-Semadeni \& Passot 1999) is not adequate to reproduce
+the flows comparable in complexity and in Reynolds numbers, and the
+situation will not change 
+in any foreseeable future. Thus only direct observational 
+studies of interstellar
+turbulence may provide us with the crucial
+information on this phenomenon. Approaching the problem one would like
+to know at least the statistics of density, velocity and magnetic field.
+In this review I briefly discuss what information emission lines can
+supply us with. I would like to quote Alyssa Goodman, who believes
+that present day technology made spectral-line
+mapping of large portions of interstellar media 
+``a booming cottage industry''. Attempts to use this wealth of observational
+data via visual inspection become fruitless and this calls for the 
+introduction of more sophisticated techniques. 
+
+
+Statistical description is a nearly indispensable strategy when
+dealing with turbulence and a big advantage of statistical techniques
+is that they extract underlying regularities of the flow and reject
+incidental details. 
+Attempts to study interstellar turbulence with statistical tools
+date as far back as the 1950s
+(see Horner 1951, Kampe de Feriet 1955, Munch 1958, 
+Wilson et al. 1959) and various directions
+of research achieved various degree of success (see reviews by
+Kaplan \& Pickelner 1970, Dickman 1985, Lazarian 1992, Armstrong, Rickett
+\& Spangler 1995). 
+Studies of turbulence statistics of ionized media were successful
+(see Spangler \& Gwinn 1990, Narayan 1992) and provided the information of
+the statistics of plasma density\footnote{Incidentally the found
+spectrum was close to a Kolmogorov one.} at scales $10^{8}$-$10^{15}$~cm. 
+This research profited
+a lot from clear understanding of processes of scintillations and scattering
+achieved by theorists (see Goodman \& Narayan 1985, Narayan \& Goodman 
+1989). At the same time 
+the intrinsic limitations of the scincillations technique
+are due to the limited number of sampling directions and difficulties
+of getting velocity information. 
+
+
+Deficiencies in the theoretical description have been, to our mind, the
+major impediments to studies of turbulence using emission lines. 
+For instance, important statistical studies of molecular clouds 
+(Dickman 1985, Dickman \& Kleiner 1985, Miesch \& Bally 1994) have not
+achieved the success parallel to that in scintillation studies. 
+
+Potentially, studies of interstellar turbulence via emission lines
+can provide statistics of turbulence in various 
+interstellar phases, including neutral gas. More importantly, velocity
+information allows one to distinguish between
+static structures and  dynamical turbulence.
+
+The difficulty of studying Doppler broadened lines stems from the fact
+that
+one has to account for both velocity and density fluctuations. Indeed,
+at any given velocity the fluctuation of emissivity may arise both
+from the actual blobs of gas moving at this velocity and from parcels
+of gas with different spatial positions but accidentally having the same
+component of velocity along the line of sight. Therefore fluctuations
+of emissivity at a given velocity would be expected even if the media were
+completely incompressible. 
+
+There exist various ways of dealing with position-position-velocity 
+(henceforth
+PPV)
+data cubes. One of them is to identify clumps and to describe their
+statistics (see Stutzki \& Gusten 1990, Williams, de Geus \& Blitz 1994).
+Another is use a more traditional set of hydrodynamic tools like
+power spectra, structure functions etc. The two statistics are interrelated
+(see Stutzki et al. 1998), but in general the relation between various
+tools is non-trivial. It seems that for answering various
+questions different statistical tools are more suitable. Therefore
+it is very encouraging that a number of techniques, including Principal
+Component Analysis (see Heyer \& Schloerb 1997) and Spectral Correlation
+Functions (Goodman 1999, Rosolowsky et al. 1999) have been 
+recently introduced to the field. 
+
+In what follows I depart from a traditional statistical hydrodynamics and
+describe how the 3D velocity and density power spectra can be extracted from
+position-position-velocity (PPV) data cubes. This choice reflects my subjective
+preferences and partially motivated by the fact that this approach relates the
+long-studied 3D density and velocity statistics (e.g. power spectra)
+with the observational data. Even with this limitation  the scope the subject
+is too broad and I shall mostly talk about atomic hydrogen (HI) studies,
+that can be viewed as a test case for the technique. I discuss advantages
+of using HI as a test case in section~2, the problem of space-velocity
+mapping in section~3 and spectra in velocity slices in section~4.
+The interpretation of 21~cm Galactic and SMC
+data is given in section~5. Possible anisotropies of statistics 
+stemming from magnetic
+field are dealt with in section~6, where a new technique for statistical 
+studies of magnetic field is suggested. I consider formation of
+emissivity enhancements that can be identified as 
+ filaments and clouds in section~7 and 
+discuss the generalization of the technique in section~8. Being aware of
+the limitations of the traditional hydrodynamic description of turbulence,
+we  describe alternative approaches, i.e.
+ 2D Genus statistics, Spectral Correlation Functions and
+Bispectrum in section~9. A short discussion of the results is given
+in section~10. 
+
+
+\section{HI as a Test Case} 
+
+Atomic hydrogen is an important component of the interstellar media 
+and many
+efforts have been devoted to its studies (see Burton 1992). 
+In terms
+of turbulence studies it has a number of advantages. For one thing, when
+dealing with HI one may in most cases disregard self-absorption.
+There are two major reasons for that: self-absorption
+is small (Braun 1997, Higgs 1999) and as shown in Lazarian (1995, henceforth
+L95), small localized absorption features typical to HI only
+marginally influence the statistics on the scales larger than their size.
+For another thing, the pervasive distribution of neutral hydrogen
+presents a sharp contrast to the localized distribution of
+molecular species, and this alleviates problems related to averaging.
+Moreover, atomic hydrogen emissivity is proportional to the
+first power of atomic density and this simplifies the analysis.
+
+HI has a substantial filling factor ($\sim 20\%$ or larger) in the Galactic
+disc and therefore its motions should reflect large scale galactic
+supersonic turbulence. At the same time, its statistics 
+may have connection with the statistics of molecular clouds. An
+additional advantage of HI is that it can be studied not
+only within our Galaxy but for the nearby galaxies as well.
+
+
+Another motivation for studies of HI statistics stems from the recent attempt
+to describe the structures in the Galactic hydrogen in order
+to estimate the fluctuations of microwave polarization arising from
+interstellar dust. This contribution is extremely important in view of
+present-day efforts in the Cosmic Microwave Background (CMB)
+ research (see Prunet \& Lazarian 1999,
+Draine \& Lazarian 1999). Some of the studies, for instance 
+one by Sethi, Prunet and Bouchet (1998),
+attempts to relate the statistics of density observed in the velocity
+space and the statistics of polarization fluctuations. If such a
+relation were possible, it would greatly alleviate the efforts to study
+polarization of cosmological origin. As an intermediate step in this work,
+however, one should relate the statistics in emissivity in the PPV
+space and density of HI in real space.
+ 
+The timing for developing statistical tools for HI studies
+is also influenced by the fact that new large data cubes, e.g.
+the Canadian Galactic Survey data (see Higgs 1999), 
+should become available soon.
+
+\section{Basic equations}
+
+\subsection{Space-Velocity Mapping}
+
+{\em Problem}\\
+The notion that the velocity fluctuation can influence emissivity
+within PPV data cubes is not new.
+Since the early-seventies 
+Butler Burton on numerous occasions claimed the importance of
+velocity fluctuations for the interpretation of 21~cm data (Burton 1970,
+1971). The
+ambiguities of inferring cloud properties from CO emission lines
+were discussed by Adler \& Roberts (1992). Using N-cloud
+simulations of spiral disks they showed that many spurious
+effects appear because of velocity blending along line of sight.
+Recently a number of researchers
+doing numerics (Pichardo et al. 1999, Vazques-Semadeni 1999) pointed out
+that pixel-to-pixel correlation between the channel maps and 
+the velocity slices of PPV data cubes tends to be larger with the
+velocity rather than the  density field.
+
+To describe power spectra of velocity and density fields, i.e. to
+express the interstellar statistics using the language that was so
+successful in hygrodynamics (Monin \& Yaglom 1972), one needs
+to disentangle velocity and denisty contributions to 21~cm emissivity
+fluctuations. 
+
+  
+{\em Approach}\\
+A quantitative 
+treatment of the effects of space-velocity mapping is given in Lazarian
+\& Pogosyan (1999, henthforth LP99). There it is assumed that the 
+velocity of a gas
+element can be presented as a sum of the regular part ${\bf v}^{reg}$
+which can arise for instance from 
+Galactic rotation, and a random, turbulent,
+part ${\bf u}$, so that
+${\bf v}^{obs} = {\bf v}^{reg} + {\bf u}$. 
+The mapping from real space to PPV coordinates corresponds to
+a transformation
+\begin{eqnarray}
+{\bf X_s} &=& {\bf X} \nonumber  \\ 
+z_s &=& A \left[ f^{-1} z - {\bf u}({\bf x}) \cdot {\bf \hat z} \right]~~~, 
+\label{eq:mapnoz}
+\end{eqnarray}
+ where henceforth we use large letters to denote vectors in the
+Position-Position plane (i.e. $xy$-plane)
+and use $z_s$ for  the velocity coordinate. The parameter 
+$A$ is just a conversion factor which specifies the units
+of $z_s$ coordinate and it is intuitively clear that this factor should
+not enter any final expressions for turbulence statistics. On the contrary,
+the shear parameter $f=\left(\delta v_z^{reg} / \delta z \right)^{-1}$
+is an important characteristic of the mapping and one expects it
+to influence our final results. For Galactic disc mapping it is convenient to choose
+$A=f$, while studies of isolated clouds correspond to a zero shear,
+i.e. $ f^{-1} \to 0$. As most work on HI has been done so far on Galactic
+disc HI, to simplify our presentation we use the former choice. With this
+definition of space-velocity mapping LP99 obtain the power spectrum
+$P_s$ in the PPV space:
+\begin{eqnarray}
+\langle \rho_s({\bf k}) \rho_s^{*}({\bf k^{\prime}}) \rangle &=& P_s({\bf k}) 
+\delta({\bf k - k^{\prime}})\nonumber \\
+P_s({\bf k}) &=& e^{-f^2
+k_z^2 v_T^2}
+\int d^3 {\bf r} \, e^{i {\bf k}  \cdot {\bf r}} 
+\Xi({\bf k}, {\bf r}),
+~~~~~~~ {\bf r}={\bf x} - {\bf x \prime}~~,
+\label{eq:roman}
+\end{eqnarray}
+where the kernel is
+\be
+\Xi({\bf k}, {\bf r})=\langle e^{i f 
+k_z (u_z({\bf x})- u_z({\bf x \prime}))}
+\rho ({\bf x}) \rho ({\bf x \prime}) \rangle~~~.
+\label{kernelXi}
+\ee
+In derivation of this expression it is explicitly assumed that the
+turbulence is statistically homogeneous in the real space coordinates
+and the average denoted by angular brackets $\langle....\rangle$
+depends only on the vector separation between points. The density Fourier
+modes in PPV space $\rho_s({\bf k})$ are uncorrelated\footnote{A treatment
+of turbulence within individual clouds is slightly different (LP99).} 
+which is reflected
+in $\delta$ function presence in the right-hand side of the first equation
+in (\ref{eq:roman}). The factor $ e^{-f^2 k_z^2 v_T^2}$, where $v_T$
+is a thermal velocity of atoms originates from averaging over thermal
+distribution and it shows that only supersonic motions are readily 
+available for statistical studies. Note, that
+expressions similar to (\ref{eq:roman})
+and (\ref{kernelXi}) were earlier obtained by Scoccimarro et al. (1999)
+in the framework of studies of Large Scale Structure of the Universe
+and this confirms the similarity of the problems studied in the two fields.
+However, the problem of ``redshift-space'' corrections to the statistics
+of galaxy distribution (Kaiser 1987)
+has been addressed either in the linear regime
+when perturbations are small  or when the velocity contribution to the
+Fourier spectrum can be factorized by a Maxwellian factor (see Hamilton 1998).
+The problem that is dealt in turbulence case is much richer as one
+has to deal with non-linear density fields transformed by coherent velocities.
+
+Note that velocity and density enter eq.~(\ref{kernelXi}) in a different
+way: velocity is in the exponent and density enters as the product
+$\rho ({\bf x}) \rho ({\bf x \prime})$. This provides an opportunity
+to disentangle the two contributions.
+
+\subsection{Spectrum in PPV Space}
+
+LP99 proves that in terms of final results
+for Lognormal distribution of density fluctuations
+and Gaussian distribution of velocity fluctuations it is safe to
+separate velocity and density in the following way
+\be
+\langle e^{i f \ldots} \rho ({\bf x}) \rho ({\bf x \prime}) \rangle =\langle e^{i f \ldots} \rangle \langle \rho ({\bf x}) \rho ({\bf x}+{\bf r}) \rangle~~~,
+\label{sep} 
+\ee
+even if density and velocity are strongly correlated. 
+It is interesting to check the degree of uncertainty that
+ the assumption (\ref{sep}) entails using numerically generated density
+and velocity fields. 
+
+For the sake of simplicity the 
+density correlation function and velocity correlation tensor
+are considered to be isotropic in Galactic 
+coordinates ($xyz$ space), i.e.
+\be
+\langle \rho ({\bf x}) \rho ({\bf x}+{\bf r}) \rangle=\xi(r)=\xi({\bf r})~~~.
+\label{xifirst}
+\ee 
+\begin{equation}
+\langle \Delta u_i \Delta u_j \rangle = \left( D_{LL}(r)-D_{NN}(r) \right) {r_i r_j \over r^2}
++D_{NN}(r) \delta_{ij}~~~,
+\label{struc}
+\end{equation}
+where $D_{LL}$, $D_{NN}$ are longitudinal
+and transverse correlation functions respectively (Monin \& Yaglom 1972), and 
+$\delta_{ik}$ equals 1 for $i=k$ and zero otherwise. These assumptions
+are not necessary, as the treatment can be provided for instance
+for axisymmetric turbulent motions (see Oughton, Radler \& Matthaeus 1997)
+as it is discussed in (L95). 
+
+The general expression for the 3D spectrum in PPV space
+is
+\begin{equation}
+P_s({\bf k}) = e^{-f^2 k_z^2 v_T^2/2}
+\int d^3 {\bf r} \, e^{i {\bf k}  \cdot {\bf r}} \xi(r)
+\exp\left[-\frac{1}{2}f^2k_z^2 D_z({\bf r})
+\right],
+\label{eq:main}
+\end{equation}
+where 
+\begin{equation}
+D_z({\bf r}) \equiv \langle \Delta u_i \Delta u_j \rangle \hat z_i 
+\hat z_j = D_{NN}(r) + [ D_{LL}(r)-D_{NN}(r)] \cos^2\theta~,
+~~~ \cos\theta \equiv {\bf \hat r \cdot \hat z}
+\label{eq:Dz}
+\ee
+is the projection of structure tensor to the $z$-axis. 
+Expression (\ref{eq:main})
+is quite general and can be used to relate arbitrary velocity and density
+statistics in galactic coordinates with the HI emissivity in the PPV space.
+
+\section{Spectra in Velocity Slices}
+  
+Observations of Galactic HI (Green 1993) revealed two dimensional spectrum
+of intensity fluctuations (see L95) and this spectrum shows power-law
+behaivior. Similar power laws for Galactic data were
+ obtained  by Crovisier \& Dickey (1983),
+Kalberla \& Mebold (1983), Kalberla \& Stenholm (1983) and for
+Small Magellanic Clouds (SMC) by Stanimirovic et al. (1999).
+Thus LP99 considered power law statistics, namely, of velocity
+${\cal P}_{3v}\sim k^{\nu}$ and density ${\cal P}_{3\rho}\sim k^{n}$,
+where ${\cal P}$ is used to denote spectra in Galactic coordinates. Note, 
+that $n<0$ and $\nu<0$ and $D_z\approx Cr^{m}$, where $m=-\nu-3$.
+Power law spectra were also reported for molecular $^{12}$CO
+(data from Heithausen \& Thaddeus 1990 and
+Falgarone et al. 1998) and $^{13}$CO (data from Heyer \& Schloerb 1997)
+lines and it looks that power law spectra are quite generic for 
+interstellar turbulence (Armstrong et al. 1997). Thus the assumption 
+of a power law statistics
+does not tangibly  constrain the range of applicability of the developed 
+theory\footnote{It is rather unnatural to expect that velocity and
+density spectra not being power laws conspire to produce power law
+emissivity.}.  
+
+For power-law spectra of density with $n>-3$ 
+the correlation functions are also power-law:
+\begin{equation}
+\xi(r)=\langle \rho \rangle^2 
+\left(1 + \left( {r_0 \over r} \right)^\gamma\right), ~~~~~ \gamma=n+3 > 0~~~.
+\label{eq:xi}
+\end{equation}
+
+Substituting Eq.~(\ref{eq:xi}) in (\ref{eq:main}) one can see that 
+\be
+P_s(|{\bf K}|,k_z)=\langle \rho \rangle^2
+\left[P_{v}(|{\bf K}|,k_z)+P_{\rho}(|{\bf K}|,k_z)\right] ~~~, 
+\label{eq:split}
+\ee
+where the part $P_v$ comes from integrating unity in Eq.~(\ref{eq:xi})
+and the part $P_{\rho}$ comes from integration the $\left( {r_0 \over r} 
+\right)^\gamma$ part. As we may see, the $P_{\rho}$ part is influenced
+by both velocity and density fluctuations, while $P_v$ part arises
+only from density fluctuations. LP99 show that an expression is 
+similar to (\ref{eq:split}) valid for
+$n<-3$.
+
+The relation between 2D spectrum in velocity
+slices and the underlying 3D emissivity spectrum in the PPV space is given
+by  
+\be
+P_2({\bf K})|_{\cal L} \sim
+{1 \over 2 \pi} \int_{-\infty}^{\infty} dk_z  P_s({\bf K}, k_z)\,
+2 \left[(1-\cos(k_z {\cal L}) ) / (k_z {\cal L})^2\right] ~~~,
+\label{eq:slice}
+\ee
+where ${\bf K}$ denotes a 2D wavevector defined in the 
+planes perpendicular
+to the line-of-sight and $|_{\cal L}$ reflects the dependence on the 
+slice thickness.
+Equation (\ref{eq:slice}) 
+represents the one dimensional integral of the three dimensional spectrum
+with the window function given by the expression in square brackets.
+It is easy to see that the thinner is 
+the velocity slice ${\cal L}$, the larger the $k_z$ range for which
+the window function is close to unity
+and therefore 
+more 3D modes contribute to 2D spectrum. 
+
+Substituting (\ref{eq:split}) into (\ref{eq:slice}) one can see that
+the two dimensional spectrum can be presented as a sum
+\begin{equation}
+P_2(|{\bf K}|)=\langle \rho \rangle^2
+\left[P_{2v}(|{\bf K}|)+P_{2\rho}(|{\bf K}|)\right]~~~,
+\label{2Dspec}
+\end{equation}
+where the expressions for $P_{2v}$ and $P_{2\rho}$ are self-evident.
+To avoid possible misunderstanding 
+I would like to stress that  $P_{2v}$ and $P_{2\rho}$
+are {\it not} 2D velocity and density spectra and, for instance,
+ $P_{2\rho}$ depends both on velocity and density statistics.
+
+Velocity fluctuations are most important for supersonic turbulence
+which is the case for cold HI. In this situation the following
+power-law asymptotics can be obtained 
+(see Table~1):
+\be
+{\rm thin~~slice}: ~~~~~~~~ C  |{\bf K}|^{-m} \gg \delta V^2 
+\label{eq:phystrans1}
+\ee
+\be
+{\rm thick~slice}: ~~~~~~~~ C  |{\bf K}|^{-m} \ll \delta V^2
+\label{eq:phystrans}
+\ee
+In other words, if the velocity dispersion $Cr^{m}$ on the 
+scale $|{\bf K}|^{-1}$
+is larger than the squared width of the channel  (in velocity units)
+the slice is termed {\it thin}. If the opposite is true  the slice is 
+termed {\it thick}. 
+
+
+
+\begin{table}[h]
+\begin{displaymath}
+\begin{array}{lrr} \hline\hline\\
+& \multicolumn{1}{c}{\rm thick~~slice}  &
+\multicolumn{1}{c}{\rm thin~~slice}
+\\[2mm] \hline \\
+P_{2\rho}({\bf K}): &
+|{\bf K}|^n;
+&|{\bf K}|^{n+m/2} \\[2mm] 
+\hline \\[3mm]
+P_{2v}({\bf K}): & |{\bf K}|^{-3-m/2}; &
+|{\bf K}|^{-3+m/2} \\[3mm] \hline
+\end{array}
+\end{displaymath}
+\caption{Asymptotics of the  components of 2D spectrum in the {\it thin}
+and {\it thick} velocity slices, $m=-\nu-3$.} 
+\label{tab:2Dspk_asymp}
+\end{table}
+
+{\it Thin Slices}\\
+It is easy to see from Table~1 that in thin slices the velocity 
+mapping makes the spectra more shallow (as $m>0$). This means that
+velocity creates a lot of small scale structures. It is also
+evident that if the $n<-3$ the $P_{2v}$ contribution dominates.
+In the opposite regime $P_{2\rho}$ contribution is important.
+
+{\it Thick Slices}\\
+The relative importance of $P_{2v}$ and $P_{2\rho}$ depends on
+whether $n>-3-m/2$ or  $n<-3-m/2$. In the former case $P_{2\rho}$
+dominates, while for the latter  $P_{2\rho}$ becomes dominant
+only when the slice is {\it very thick}, i.e. a substantial portion
+of the line is integrated over. Indeed, it is easy to see
+that integration over the line-width leaves only density information.
+ In the intermediate situation if the density spectrum is steep, i.e.
+$n<-3-m/2$
+ $P_{2v}$ provides most of the contribution to $P_2$.
+
+For warm HI the thermal velocity dispersion is comparable with
+the turbulent one. Thus fluctuations of intensity arise mostly
+from density inhomogeneities and the analysis in L95
+is applicable. The amplidude of fluctuations arising from the
+warm phase of HI is suppressed due thermal velocity smearing
+effects. Therefore in the mixture of the warm and cold phase
+the contribution of the cold phase to the measured spectrum
+is likely to dominate (LP99).
+
+
+\section{Statistics of Diffuse HI}
+
+\subsection{Analysis of data}
+
+One of the most thorough jobs of obtaining 21~cm statistics was done
+by Green (1993). His observations
+of the HI emission were accomplished with the Synthesis
+Telescope of the Dominion Radio Astrophysical Observatory (DRAO)
+towards $l= 140^{\circ}, b=0^{\circ}$ ($03^{h} 03^{m} 23^{s},
++58^{\circ} 06' 20'$, epoch 1950.0) and they revealed a power law spectrum of
+2D intensity. This spectrum is proportional to $P_2({\bf K})|_{\cal L}$
+and its interpretation depends on whether the slicing is thick or thin.
+To answer this question one has to estimate the dispersion of turbulent
+velocity at the scales under study and compare it to the velocity
+thickness of the slice (see eq.(\ref{eq:phystrans1}) and (\ref{eq:phystrans}).
+Assuming that velocity variations at the scale $30$~pc amount
+to $10$~km/s and arise from the Kolmogorov turbulence, the structure
+functions of velocity are
+\be
+D_{LL}(r)\approx 100
+\left(\frac{r}{30 {\rm pc}}\right)^{2/3}~{\rm km}^2{\rm s}^{-2}~~~,
+\ee
+The width of the interferometer channels combined to give
+a single data point in Green's dataset is $\delta V =5.94~\mathrm{km/s}$.
+The slice thickness in parsec is ${\cal L}\approx \delta V f$~pc, 
+and varies from
+$\approx 600$~pc for the closest slices to  $\approx 2200$~pc for the 
+distant ones\footnote{Note, that the cut-off due to thermal velocity (see
+section~2.1) in Warm Neutral Medium (see a table of idealized phases
+in Draine \& Lazarian (1999) ) is $\sim 6$~km/s. If the WNM constitutes
+the dominant fraction of the neutral phase (Dickey 1995) then the
+velocity resolution above is optimal and no further decrease in $\delta V$
+will result in getting new information. However, if close to Galactic
+plane Cold Neutral Media constitutes a substantial portion of mass, the
+increase of velocity resolution up to 1~km/s is desirable.}. The wavenumber of transition from {\it thin} to {\it think}
+slice given by eq.~(\ref{eq:phystrans}) is equal $0.16\,pc^{-1}$. 
+
+In figure~1 the turbulence scales covered by Green's study are shown.
+The smallest 
+$|\bf K|$ span from $\sim 1/3$~pc$^{-1}$ for the closest slices
+to $1/200$~pc$^{-1}$ for the distant ones.
+It is obvious that most of the measurements correspond to the
+thin slice regime. 
+
+\begin{figure}[ht]
+\centerline{\psfig{file=flat_rot.eps,width=7cm,angle=0}}
+%\centerline{\epsfxsize=3.in\epsfbox{flat_rot.eps}}
+\caption{The variations of geometric scales with the sampling velocity are
+shown (from LP99). The upper curve corresponds to the variations of 
+the correlation
+scale $\lambda=(f^2C)^{1/(2-m)}$ 
+in the velocity space. Physically this is the scale at which the
+square root of velocity dispersion $\sim (C\lambda^m)^{1/2}$ become
+equal to the difference of the regular velocities due to Galactic rotation.
+The middle curve corresponds to the variations
+of the slice thickness $L$. The darkened area in the Figure depicts
+to the range of the turbulence scales under study in Green (1993). 
+The solid curve within the darkened area corresponds to the interferometric
+measurements with the baseline 21~m. The lower
+horizontal line denotes value of
+$|{\bf K}|^{-1}$ which separates {\it thin} (above) and {\it thick} slice
+regimes (below) for cold HI. The contribution of fluctuations from the
+warm phase is suppressed when the slices are thin for cold HI and thick
+for warm HI (LP99).}
+\label{fig:flat_rot}
+\end{figure}
+ 
+As we mentioned earlier, whether $P_{2v}$ or $P_{2\rho}$ dominates
+the observed emissivity spectra in the thin slice regime
+depends on whether $n$ is larger or smaller than $-3$. If $n>-3$, 
+ $P_{2\rho}$ is dominant and the observations by Green (1993) reveal 
+the spectrum with index $n+m/2$. For $m=2/3$ the spectrum of emissivity
+obtained by Green (1993), namely the emissivity with the index $\sim -2.7$,
+corresponds to $n \sim -3$. If, however, the density spectrum is 
+steep (i.e. $n<-3$), the
+fluctuations of 21~cm intensity observed by Green (1993) can arise
+from velocity fluctuations. In this case the spectral index is
+$-3+m/2$. For $m=2/3$ one gets the slope $-8/3\approx -2.7$ which
+is exactly what is observed. The question now is whether nature
+conspires to create the density spectrum with $n\approx -3$ and
+thus make the slopes of  $P_{2v}$ or $P_{2\rho}$ identical in
+thin slices or we observe  $P_{2\rho}$, while $n<-3$.
+
+To answer this question one has to consider thick slices of data.
+Unfortunately, for thick slices of data one has to account for
+ lines of sight being not parallel. This problem was studied
+by Lazarian (1994) for the case of density statistics, but the
+study has not been generalized so far for the case when both density
+and velocity contribute to emissivity. Thus external galaxies provide
+a better case for study thick slices.  Data for Small
+Magellanic Cloud
+ (SMC) in Stanimirovic et al. (1999) shows the
+steepening of the slope from $\approx -3$ for the slices obtained
+with the maximal velocity resolution  to $n\approx-3.5$ for data 
+integrated over
+the entire emission line
+(Stanimirovic, private
+communication) which corresponds to our theoretical predictions for
+long-wave dominated density spectrum with the index $n \approx -3.5$. 
+Thus the set of Green's and Stanimirovic's data is consistent with
+an interpretation that both velocity and density exhibit spectra
+close to Kolmogorov.
+
+A potential difficulty of ``Kolmogorov'' cascade interpretation is
+that the SMC data show power-law slope up to 4~kpc scale. To explain
+Doppler broaderning of molecular lines one has to accept that the
+energy is being injected at large scales. However scales of
+several kpc look excessive.
+Injection
+of energy at such large scales is possible in the form of superbubles,
+but the details and the very possibility of the cascade in these circumstances
+is unclear. 
+
+
+
+
+\subsection{Further tests}
+
+The ``Kolmogorov'' interpretation above apparently needs further testings.
+There are various pieces of evidence that could be
+interpreted in favor of the
+spectrum of density being shallow, i.e. $ \approx -3$. 
+However, our analysis shows that this interpretation is not substantiated.
+
+For instance, Braun (1999) reports a power law
+index of $-3$ for the spectrum of 21~cm emission
+from structures near the North-East major axis of M31 galaxy. 
+However, he uses not the whole spectral lines, as we do in our
+treatment but the maximal values of velocity only 
+(Braun 1999, private communication). The interpretation
+of this result in terms of the power spectra as discussed in L95 and
+LP99 is impossible, as the treatment of data is very different.
+
+Shallow spectrum of Far Infrared emission
+from dust (Wall \& Waller 1998, Waller et al 1998) does not support the
+the shallow HI density spectrum either. According to Stanimirovic (1999,
+private communication) the shallow spectrum of Far Infrared emission
+when converted into dust column density provides a steep ``Kolmogorov''-type
+spectrum. 
+
+One can argue that a 
+possible hint in favor of HI density being short wave dominated comes from
+molecular data discussed in Stutzki et al. (1998). There  for both
+ $^{12}$CO (data from Heithausen \& Thaddeus (1990) and
+Falgarone et al. (1998))
+and $^{13}$CO (data from Heyer \& Schloerb (1997))
+transitions the spectrum of intensity was observed
+to have a power law index $\sim -2.8$. As the data is averaged over
+velocity, the fluctuations of intensity are due to density fluctuations
+and the spectrum of density should have the same
+slope as the spectrum of emissivity (L95), provided
+that the transitions are optically thin. The problem is that they are
+not thin and therefore the analysis above is not applicable.
+
+Attempting to establish the actual underlying spectrum one may 
+use one dimensional spectrum introduced in LP99
+\begin{equation}
+P_1(k_z) = \int d{\bf K} P_s({\bf K}, k_z)
+\label{16}
+\end{equation}
+Similar to two dimensional spectrum $P_1$ can be presented  as a sum of
+$P_{1\rho}$, which scales as $k_z^{2(n+2)/m}$ and $P_{1v}$, which
+scales as $k_z^{-2/m}$. Naturally, if $n<-3$ $P_{1v}$ dominates,
+while $P_{1\rho}$ dominates for $n>-3$. The analysis of data using 
+$P_1(k_z)$ has not been done so far.
+
+
+
+\section{Anisotropies and Magnetic field}
+
+\subsection{Goldreich-Shridhar Turbulence}
+
+It is natural to expect that dynamically important magnetic field makes
+interstellar turbulence anisotropic (Montgomery 1982, Higton 1984). Indeed, 
+it gets difficult for
+hydrodynamic motions to bend magnetic fields at small scales if
+the energy density of the magnetic field and hydrodynamic motions 
+are comparable at large scales.
+The turbulence in ionized gas
+has been found to be anisotropic and its  Kolmogorov-type spectrum of 
+plasma density fluctuations
+observed via radio scintillations and scattering (see Armstrong 
+et al. 1995 and references therein) has been interpreted recently 
+as a consequence of
+a new type of MHD cascade by Goldreich \& Sridhar (1995).
+ The Goldreich-Shridhar model of turbulence\footnote{A qualitative
+discussion of the model and the role of reconnection for the
+cascade can be found in Lazarian \& Vishniac (1999).} differs considerably
+from the Kraichnan one (Iroshnikov 1963, Kraichnan 1964).
+It accounts for the fact that hydrodynamic motions can easily mix up
+magnetic field lines in the plane perpendicular to the direction of
+the mean field (see discussion in Lazarian \& Vishniac 1999). 
+Such motions provide eddies elongated in the 
+field direction and the velocity spectrum close to the Kolmogorov one.
+
+The Goldreich-Shridhar turbulence is anisotropic with eddies stretched 
+along magnetic
+field. 
+The wavevector component parallel to magnetic field $k_{\|}$
+scales as $k_{\bot}^{2/3}$, where $k_{\bot}$ is a wavevector component 
+perpendicular
+to the field. Thus the degree of anisotropy increases with the
+decrease of  scale. 
+
+\subsection{Anisotropies and magnetic field direction}
+
+It is both  challenging and important to determine
+the degree of anisotropy for the HI statistics for various parts of
+the Galaxy. This information can provide an insight
+to the nature of HI turbulence and may be used as a diagnostic for the
+interstellar magnetic field. For instance, measurements  of
+ the structure functions
+of HI intensity 
+\be
+S(\theta,\phi)=\langle (I({\bf e}_1)-I({\bf e}_2))^2\rangle
+\ee 
+as a function of a positional angle $\phi$ for individual
+subsets of data
+should reveal magnetic field direction in various portions of the sky, 
+if  the turbulence is anisotropic as we expect it to be. This technique 
+is somewhat analogous to a technique of finding magnetic field direction
+using the fluctuations of synchrotron radiation (see Lazarian 1992) but
+its applicability may be much wider. 
+
+So far, the attempts to measure anisotropy in HI are limited to the 
+Green (1994) study,
+where no anisotropy was detected. Apparently a better analysis is
+needed. For the slices with high degree of
+anisotropy the statistical technique can be improved as suggested in L95. 
+
+\section{Structures in PPV space}
+
+PPV data cubes, e.g. HI data cubes, exhibit a lot of small scale 
+emissivity structure\footnote{It was noticed by Langer, 
+Wilson \& Anderson (1993) that more structure
+is seen in PPV space than in the integrated intensity maps.}. The question
+is what part of them is real, i.e. is associated with density enhancements
+in galactic coordinates and what part of them is produced by velocity 
+fluctuations. A related question is whether the structures we see
+ are produced dynamically, through
+forces, e.g. self-gravity, acting on the media or they may be produced
+statistically exhibiting the properties of random field. The second
+question was partially answered in Lazarian \& Pogosyan (1997), where it 
+was shown that density fluctuations with Gaussian distribution and
+power spectra result in filamentary structures. The structures become
+anisotropic and directed towards the observer when the velocity effects
+are accounted for\footnote{There is a distant analogy between this effect
+and the ``fingers of God'' effect (see Peebles 1971) in the
+studies of large scale structure of the Universe.} (see Lazarian 1999, fig.~2). 
+
+The issue of density enhancements produced by velocity fluctuations is
+closely related to the statistics of ``clouds'' observed in
+PPV space. The results on velocity mapping that we discussed earlier
+suggest that spectra of fluctuations observed in PPV velocity slices
+are more shallow than the underlying spectra. This means
+more power on small scales 
+or, in other words, more small scale structure (``clouds'')  
+appears in the PPV slices due to velocity fluctuations.
+
+It has been believed for decades that emission cloud surveys 
+(see Casoli et al. 1984, Sanders et al. 1985, Brand \& Wouterloot 1995)
+provide a
+better handle on the actual spectrum of cloud mass and sizes than
+the extinction surveys (see Scalo 1985) because
+the velocity resolution is available. These two sorts of survey 
+present different slopes for clump sizes and the difference cannot
+be accounted through occlusion of small clouds in the extinction surveys
+by larger ones (Scalo \& Lazarian 1996). In view of the discussion above
+it looks that extinction survey may be closer to the truth, while a lot of
+structure detected via analyzing PPV cubes is due to velocity caustics.
+Paradoxically enough, emission data integrated over the spectral lines
+may provide a better handle on the distribution of cloud sizes compared
+to high resolution spectral-line data cubes. Averaging over velocity
+ results in the distortions of the cloud size spectra due to occlusion
+effects, but these effects can be 
+accounted for using the formulae from Scalo \& Lazarian (1996).        
+
+\section{Generalization of the technique}
+
+The formalism was described above in terms of HI power law statistics.
+It is obvious that it can be modified to deal with arbitrary statistics
+and with a variety of emission transitions. Here we briefly discuss
+complications which a generalization of the technique in order to be
+applicable to molecular clouds and ionized media may encounter. 
+
+\subsection{Forward and inverse problems}
+
+A considerable number of researchers believes that self-similar behavior 
+reflected in power law statistics is a characteristic feature of the
+interstellar turbulence including the molecular cloud turbulence
+(see Elmegreen \& Falgarone 1996). However, 
+some researchers (e.g. Williams 1999) see departures from a
+power law, e.g. signatures of a characteristic scale. In those cases,
+one still can find the underlying density and velocity 
+statistics solving {\it forward} and {\it inverse} problems (see
+Lazarian 1999).
+
+To solve the forward problem one needs to use expressions for the observable
+statistics, e.g. expressions for 2D and 1D spectra in PPV space (see
+eqs~(\ref{eq:main}) 
+and (\ref{16})) and fit the observable statistics varying the input of the
+velocity and density statistics. Naturally, the question of uniqueness
+for such solutions emerge, but with a reasonable choice of input 
+parameters one may hope to avoid ambiguities. 
+
+A different approach involves the inversion of input data. 
+Inversion also requires a model, but for the case of turbulence
+studies the model can be quite general and usually includes
+some symmetry assumptions, like the assumption of isotropy or
+axial symmetry of turbulence statistics (Lazarian 1994a, L95). For the
+case of turbulence the inversion has been developed for statistics
+of density (L95, Lazarian 1993) and magnetic field (Lazarian 1992, 
+Lazarian 1994a). A remarkable property of
+the inversion for turbulence statistics is that it allows analytical
+solutions, which shows that the inversion is a well posed procedure 
+in the mathematical sense. The procedure for inverting velocity
+data should be analogous to inverting density \& magnetic field
+statistics, but has not been developed so far.
+We expect  wide application of forward and especially
+inverse problem technique
+when the deviations from self-similar behavior become apparent in the
+data. 
+
+\subsection{Various Transitions}
+
+As we discussed earlier, one of the advantages of using HI as a test
+case is that the emissivity is proportional to the column density. 
+This is true for some optically thin transitions in molecular clouds,
+but fails when absorption is important. My analysis showed that
+the absorption is relatively easy to account for if it arises from
+dust, but much harder to deal with if it is self-adsorption. In the
+former case most of the analysis above is valid provided that the
+turbulence scale under study is much smaller than the extinction
+length.
+
+Homogeneity of sample is another major concern for studies of
+turbulence in molecular clouds (see Houlahan \& Scalo 1990).
+Filtering the data (see Miesch \& Bally 1994), application of
+wavelets (see Stutzki et al. 1998) or both are required. However, it seems
+that as the resolution of data improves the effects of cloud edges
+get less important and easier to take care of.
+
+Some emissivities, e.g. those of H$\alpha$ lines are proportional
+to the squared density of species. However, it is possible to generalize 
+the technique above 
+for those transitions and provide a quantitative treatment of
+turbulence in ionized
+emitting media, e.g. of HII regions (O'Dell 1986, O'Dell \& Castaneda 1987).
+ 
+
+ 
+
+  
+
+
+
+\section{Beyond Power Spectra}
+
+The approach that we discussed so far can be characterized as an
+interpretation of the emissivity spectra\footnote{
+In fact we do not distinguish between spectra and correlation functions.
+The two statistics are related via Fourier transform and provide
+an equivalent description. In some particular circumstances one or the
+other may be more convenient, however.} in terms of the underlying
+statistics of velocity and density.
+ 
+
+The advantage of this approach is that no numerical 
+inversion (see Lazarian 1999) is performed and thus one should not
+worry about increase of the data noise. The power spectra are
+widely used in hydrodynamics 
+and therefore there is hope to relate the  statistics
+of interstellar turbulence with something simple and
+more familiar like 
+Kolmogorov-type cascade.
+
+In spite of all these advantages, the information that this 
+approach can supply us with is limited. Indeed, media clustered by
+self-gravity and more diffuse media may have the same index of
+power spectrum, while being very different. In general, 
+statistical measures borrowed from hydrodynamics may not
+be adequate while dealing with
+interstellar
+media. Indeed, we have to address particular questions,
+e.g. the question the identification of star-forming regions,
+which are beyond the standard hydrodynamic description.
+Therefore
+attempts to introduce new descriptors are worthy of high praise.
+It may happen that in answering specific questions one has to
+use particular descriptors. 
+
+\subsection{Spectral Correlation Function}
+
+A new tool termed ``spectral correlation function''  or SCF has been
+recently introduced to the field (Goodman 1999, Rosolowsky et al. 1999). 
+It compares neighboring spectra with each other. For this
+purpose the following measure is proposed:
+\be
+S(T_1, T_0)_{s,l}\equiv 1-\left(\frac{D(T_1,T_2)}{s^2\int T_1^2(v)dv+
+\int T_0^2(v)dv}\right)~~~,
+\ee
+where the function
+\be
+D(T_1,T_2)_{s,l}\equiv \left\{\int [s T_1(v+l)-T_0(v)]^2dv\right\}
+\ee
+and the parameters $s$ and $l$ can be adjusted. One way to choose
+them is to minimize $D$ function. In this case $S$ function
+gets sensitive to similarities in the shape of two profiles. 
+Fixing $l$, $s$ or both
+parameters one can get another 3 function that are also 
+sensitive to similarities in amplitude, velocity offset
+or to both parameters.
+
+The purpose of those functions is to distinguish regions with various
+star forming activity and to compare numerical models with observations.
+To do this histograms of SCF are compared with
+histograms of SCF obtained for the randomized spatial positions. 
+This allows to models to be distinguished on the basis of their clustering
+properties. First
+results reported by Rosolowsky et al. (1999) are very encouraging. 
+It was possible to find differences for simulations corresponding to
+magnetized and unmagnetized media and for those data sets for which
+an earlier analysis by Falgarone et al. (1994) could not find the difference. 
+The mathematical development of this new tool is under way (Padoan et al. 2000)
+and we expect new exciting results to be obtained in the
+near future.
+
+A few comments about spectral correlation functions may be relevant.
+First of all, by its definition it is a very flexible tool. In the
+analysis of Rosolowsky et al. (1999) the SCF were calculated for the
+subcubes over which the original data was divided. In this way SCF
+preserves the spatial information and in some sense is similar
+to cloud-finding algorithms (see Stutzki \& Gusten 1990, Williams,
+de Geus \& Blitz 1994). However, one may fix the angular separation between
+the studied spectra and then the technique will be more similar to
+the traditional correlation function analysis that is sensitive to
+turbulence scale rather than to positional information. I believe that
+this avenue should be explored in future along with other more sophisticated
+techniques that can be applied to SCF.  
+At first glance, it looks counterproductive to get a whole lot of 
+various correlations using SCF as the input data. However,
+we must find a way of distinguishing regions with various physical
+properties and we are still in search for the best descriptors.
+
+At the moment the distinction between various interstellar regions and
+the sets of simulated data is made by eye examining the histograms of
+SCF. With more information available it seems feasible to use wavelets
+that will emphasize some characteristics of the histograms in order
+to make the distinction quantitative. Construction these wavelets
+will be the way of ``teaching'' SCF to  extract 
+features that distinguish various sets of data.   
+
+\subsection{Genus Statistics}
+
+The topology of ISM is an essential characteristic
+of the media. Genus analysis has been proved to be a useful tool for
+characterizing topology of the Universe (see Gott et al. 1989) and
+therefore it is tempting to apply it to the ISM studies. 
+
+The two dimensional genus analysis can be directly related to the
+media topology. By 2D maps we mean here maps integrated over the
+emission line, i.e. total intensity maps.
+
+A two-dimensional genus is (Melott et al. 1989)
+$$ G_2(\nu_t)=({\rm number~of~ isolated~ high~ density~ regions}) - 
+({\rm number~ of~ isolated~ low~ density~ regions})$$
+where $\nu_t$ denotes the dependence of genus on the threshold density in units
+of standard deviations from the mean. It is obvious that if one raises
+the density threshold from the mean,
+the low density regions coalesce and the genus
+becomes more positive. The opposite is true if $\nu_t$ decreases. Thus
+for Gaussian fluctuations one expects genus to be antisymmetric about
+zero, but the actual distributions should be able to
+ reveal ``bubble'' or ``meatball''
+topology of various parts of the 
+ISM. Algorithms exist for calculating genus for
+2D maps, e.g. microwave background maps (Colley et al. 1996) and
+therefore the application of genus statistics to interstellar 
+maps is straightforward (and long overdue).
+
+The 3D genus statistics (see Gott et al. 1989)
+in PPV space is less easy to interpret. As we
+discussed earlier, a lot of structures there are due to velocity caustics
+and the relation of the structures in galactic coordinates and PPV space
+is not obvious. However, it seems interesting to apply genus analysis
+to the PPV space in search for another statistical tool to distinguish
+various interstellar regions. After all, SCF introduced by Alyssa Goodman
+do not have a straightforward relation to the known parameters, 
+but are very useful.
+
+
+\subsection{Bispectrum}
+
+Attempts to use multipoint statistics are a more traditional way to
+remove the constraints that the use of two point statistics, e.g.
+power spectra entails. Unfortunately, very high quality data is
+needed to obtain the multipoint statistics. 
+
+Among multipoint statistics, bispectrum (see Scoccimarro 1997)
+seems the most promising. This is partially because it has been successfully
+used in the studies of the Large Scale Structure of the Universe.
+
+Bispectrum is a Fourier transform of the three point correlation
+function and if the power spectrum $P({\bf k})$ is defined as
+\be
+\langle \delta\rho({\bf k}_1)\delta 
+\rho({\bf k}_2)\rangle = P({\bf k})\delta_D ({\bf k}_1+{\bf k}_2)
+\ee
+where $\delta_D$ is the Dirac delta function that is zero apart from
+the case when ${\bf k}_1+{\bf k}_2=0$, the bispectrum $B_{123}$
+is
+\be
+\langle \delta\rho({\bf k}_1)\delta \rho({\bf k}_2)\delta \rho({\bf k}_3)
+\rangle=B_{123}\delta_D ({\bf k}_1+{\bf k}_2+{\bf k}_3)   
+\ee
+
+It is advantageous to use 
+``hierarchical amplitude'' (Fry \& Seldner 1982) statistics
+\be
+\Phi_{123}\equiv \frac{B_{123}}{P({\bf k}_1)P({\bf k}_2)+ P({\bf k}_2)
+P({\bf k}_3)+P({\bf k}_3) P({\bf k}_1)}
+\ee
+which is for power law spectra is a scale independent quantity.
+
+In the studies of Large Scale Structure the hierarchical amplitudes
+were calculated for various initial conditions to compare with observations.
+In interstellar media it is advisable to compare various regions of
+sky using the tool. Impediments for the use of the technique stem from
+the increase of noise with the use of multipoint statistics and the
+problems of averaging along the line of sight. 
+The problems should be addressed in the future.
+
+\subsection{Other techniques}
+
+A wavelet technique (see Gill \& Henriksen 1990, Langer, Wilson \&
+Anderson 1993, Rauzy, Lachieze-Rey \& Henkiksen 1993) 
+is discussed in this volume by Stutzki who
+proposed a so called $\Delta$- variance technique (Stutzki et al. 1998)
+which is is related to wavelet transforms (Zielinsky \& Stutzki 1999).
+Wavelets potentially are a versatile tool that can filter out the
+large scale inhomogeneities of the data and concentrate the analysis
+on the scales of interest (see Stutzki, this volume).
+
+Another useful statistical tool is the Principal Component Analysis
+(PCA). This tool was employed to spectral line imaging studies  of 
+the interstellar medium by Heyer \& Schloerb (1997). The goal of the
+PCA is to determine a set of orthogonal ``axes'' $u_{kl}$ for which the 
+the variance of the data is maximized. In the case of the data in
+$n$ points with $p$ velocity (spectrometer) channels for each point
+the data can be presented as $\delta T_{ij}=T_{ij}-\langle T_{ij} \rangle_n$, 
+where $T_{ij}$ is the temperature at the channel $j$ at a position $i$
+and $\langle ... \rangle_n$ denote averaging over positions. Maximazing
+the variance means maximizing the expression $y_{ij} y_{ij}=u_{ik}S_{jk}
+u_{ij}$, where summation over the repeating indexes is implied and
+$S_{ik}=\langle \delta T_{ij} \delta T_{jk} \rangle_n$. In practice finding of
+$u_{ij}$ amounts to solving a set of eigenvalue equations 
+$S_{ik}u_{kj}=\lambda u_{ij}$. To visualize the variance related to $l$-th 
+principal component eigenimages are constructed from the projections of 
+$T_{ij}$ onto the eigenvector, i.e. $l$th eigenimage at point
+ $(r_i)$ is $\delta T_{ij}u_{l,j}$.  
+Heyer \& Schloerb (1997) showed that using PCA technique it is possible
+to decompose large-scale spectroscopic images of molecular clouds. Their
+analysis enabled them to calculate the velocity - scale relations for
+a number of cloud complexes. In terms of the statistical analysis
+presented above, PCA provides a means of filtering out large scale
+features responsible for the largest contribution to the global variance.
+This makes the sample more homogeneous and suitable for describing using
+correlation functions and power spectra. The potential of this 
+statistical tool is to be further explored. It is likely that combining
+the various set of data (for instance, HI and CO) more interesting
+correlations can be obtained via PCA technique.
+
+\section{Discussion}
+
+It is not possible to cover all the various interesting approaches
+that have been tried in order to study interstellar turbulence via emission
+lines. For instance, we omitted a discussion of
+ 3D correlation functions in PPV space  introduced
+in Perault et al. (1986). We did not cover studies of turbulence
+using centroids of velocity (see Dickman 1985) either.
+One reason for this is that I believe that the statistics of
+velocity centroids have to be described in terms of underlying velocity
+and density.
+
+
+A search for tools to deal with the interstellar turbulence has been
+intensified recently and this shows 
+that there is deep understanding
+in the community that the wealth of observational data must be explored
+and it is essential to compare observations and numerical simulations.
+I personally believe that the development of theoretical approaches
+to dealing with data has become at this point not less important than
+obtaining the data.
+
+Most of the present review I devoted to dealing with power spectra which
+reflects my personal preferences. Although far from being unambiguous,
+the power spectra were most intensively studied in hydrodynamics and
+the MHD theory and therefore they provide a bridge between  idealizations
+that we partially
+understand and {\it terra incognita} of  interstellar turbulence.
+Whether this approach is useful for a particular phase of the 
+interstellar media is not clear {\it a priori}. We may or may not have
+any self-similarity indicating a turbulent cascade. However, at least
+for HI it seems that the approach is promising. Indeed, we managed
+to relate, although tentatively, the statistics of 21~cm emission with
+the statistics of Goldreich-Shridhar cascades. The next class of objects
+ to study using the technique should be molecular clouds.
+
+Although studies of molecular clouds are expected to face more problems,
+some of them mentioned earlier on, it is likely that the underlying 3D
+statistics will be soon obtained for the optically thin molecular lines.
+Comparison of this statistics with that in diffuse media should provide
+an insight to the nature of interstellar turbulent cascade and 
+turbulent support of molecular clouds.
+
+However, the limitations of the power-spectrum approach make it necessary
+to use alternative tools such as wavelets, genus statistics, principal
+component analysis and develop new ones such as spectral correlation function
+even though the relation between their output and the familiar
+notions from hydrodynamics is not always clear. In studies of
+interstellar medium one has  to address particular questions, e.g.
+the question of star formation and therefore appearance of specialized
+tools is only natural.
+
+\section*{Summary}
+
+~~~~1. Velocity and density power spectra can be obtained
+from observed emissivities. Velocity fluctuations make emissivity 
+spectra in velocity slices shallower.
+This results in much of small scale structure in PPV space
+that can be erroneously interpreted as interstellar clouds or clumps. 
+
+2. Turbulence is likely to be anisotropic with magnetic field defining
+the anisotropy direction. This should allow a new way of studying 
+magnetic field.
+
+3. The available wealth of observational data motivates the development
+of new tools for data handling.    
+
+
+\section*{Acknowledgments}
+
+This review is partially based on the results that I obtained together
+with Dmitry Pogosyan. Discussions with Chris McKee, 
+John Scalo, Steve Shore and Enrique Vazquez-Semadeni
+are acknowledged. I am grateful to Robert Braun and Snezana Stanimirovic
+for helpful input on data reduction procedures. 
+
+%\section*{Appendix}
+
+
+\section*{References}
+
+
+\begin{thebibliography}{99}
+\bibitem{} Adler, D. S. \& Roberts, W. W. Jr. \Journal{1992}{\ApJ}{384}{95}
+\bibitem{} Armstrong, J.M., Rickett, B.J., \& Spangler, S.R.
+\Journal{1995}{\ApJ}{443}{209}
+\bibitem{} Brand, P. \& Wouterloot, J.G.A. \Journal{1995}{\AaAp}{303}{851}
+\bibitem{} Braun, R. \Journal{1997}{\ApJ}{484}{637}
+\bibitem{} Braun, R. (1999) in {\em Interstellar Turbulence}, eds. 
+Jose Franco \& 
+Alberto Carraminana, CUP, p. 12
+\bibitem{}  Burton, W.B. \Journal{1970}{\em A\&AS}{10}{76}
+\bibitem{}  Burton, W.B. \Journal{1971}{\AaAp}{10}{76}
+\bibitem{}  Burton, W.B. (1992) in  {\em Galactic Interstellar
+Medium}, eds. Pfenninger~D., Bartholdi~P., Springer-Verlag, p.~1 
+\bibitem{} Casoli, F., Combes, F., \& Gerin, M. \Journal{1984}{\AaAp}{133}{99}
+\bibitem{} Colley, W.N., Gott, J.R. \& Park, C. \Journal{1996}{\MNRAS}
+{281}{L82} 
+\bibitem{} Dickey, J.M.,  (1995),  in {\em The Physics of the 
+Interstellar Medium
+and Intergalactic Medium},  eds Ferrara~A., McKee~C.F., Heiles~C.,
+Schapiro~P.R., ASP Conf. Ser. V80.
+\bibitem{} Dickman, R.L.,  (1985),  
+in  {\em Protostars and Planets II}, eds Black~D.C. and Mathews~M.S., 
+Tucson:  University of Arizona, 150
+\bibitem{}   Dickman, R.L. \& Kleiner, S.C. \Journal{1985}{\ApJ}{295}{479}
+\bibitem{} Draine B.T. \& Lazarian, A. (1999), in {\em Sloan Summit 
+on Microwave 
+Foregrounds}, ed. A. de Oliveira-Costa and M. Tegmark, in press
+\bibitem{} Elmegreen, B.G. \& Falgarone, E. \Journal{1996}{\ApJ}{471}{816}
+\bibitem{} Falgarone, E., J.-L. Puget, \& Perault, M. \Journal{1992}{\AaAp}
+{257}{715}
+\bibitem{} Falgarone, E. et al.  \Journal{1998}{\AaAp}{331}{669}
+\bibitem{} Falgarone, E., Lis, D. C.,
+ Phillips, T. G., Pouquet, A.;
+ Porter, D. H., Woodward, P. R. \Journal{1994}{\ApJ}{436}{728}
+\bibitem{}  Gautier, T.N., Boulanger, 
+F., P\'{e}rault, M. \& Puget, J.-L. \Journal{1992}{\AJ}{103}{1313}
+\bibitem{} Gill, A.G. \& Henriksen, R.N. \Journal{1990}{\ApJ}{365}{L27}
+\bibitem{} Goldreich, P. \& Sridhar, S. \Journal{1995}{\ApJ}{438}{763}
+\bibitem{} Goodman, A.A. 1999, 
+http://cfa-www.harvard.edu/~agoodman/scf/SCF/scfmain.html
+\bibitem{} Goodman, J., \& Narayan, R. \Journal{1985}{\MNRAS}{214}{519}
+\bibitem{} Gott, J.R., Mao, S., \& Park, C. \Journal{1992}{\ApJ}{385}{26}
+\bibitem{} Gott, J.R., Melott, A.L., \& Dickinson, M. \Journal{1986}{\ApJ}{306}
+{341}
+\bibitem{} Gott, J.R., Park, C., Juskiewicz, R., Bies, W.E., Bennett,
+D.P., Bouchet, F.R., Stebbins, A. \Journal{1990}{\ApJ}{352}{1}
+\bibitem{} Gredel, R. (1999) in 
+{\em Interstellar Turbulence}, eds. Jose Franco \& Alberto Carraminana, CUP,
+140
+\bibitem{} Green, D.A. \Journal{1993}{\MNRAS}{262}{328}
+\bibitem{} Green, D.A. \Journal{1994}{\em Ap\&SS}{216}{61}
+\bibitem{} Hamilton, A.J.S. 1998, in {\em The Evolving Eniverse}, ed.
+D. Hamilton, Dordrecht: Kluwer, p. 185
+\bibitem{} Heithausen, A., \& Thaddeus, P. \Journal{1990}{\ApJ}{353}{L49}
+\bibitem{} Henriksen, R.N. \Journal{1994}{\em Ap\&SS}{221}{25}
+\bibitem{} Heyer, M.H., \& Schloerb, F.P. \Journal{1997}{\ApJ}{475}{173}
+\bibitem{} Higgs, L.A. (1999), in {\em New Perspectives on
+the Interstellar Medium}, ASP 168, p. 15
+\bibitem{} Higdon, J.C. \Journal{1984}{\ApJ}{285}{109}
+\bibitem{} von Horner, S. \Journal{1951}{\em Zs.F. Ap.}{30}{17}
+\bibitem{} Houlahan, P., Scalo J., \Journal{1990}{\ApJ}{72}{133}
+\bibitem{} Iroshnikov, P.S. \Journal{1963}, {\em AZh} {40} {742}
+\bibitem{} Kaiser, N. \Journal{1987}{\MNRAS}{227}{1}
+\bibitem{} Kalberla, P.M.W. \& Mebold, U. \Journal{1983} 
+{\em Mitt Astr. Ges.}{58}{101}
+\bibitem{} Kalberla, P.M.W. \& Stenholm, L.G. \Journal{1983}
+{\it Mitt. Astron. Ges.} {60} {397}
+\bibitem{} Kamp\'{e} de F\'{e}riet, J. (1955), 
+in: {\em Gas Dynamics of Cosmic Clouds}, Amsterdam: North-Holland, 134
+\bibitem{}  Kaplan, S.A. \& Pickelner, S.B. (1970), 
+ {\em The Interstellar Medium}, (Harvard University Press)
+\bibitem{}  Kerr, F.J. \& Lynden-Bell, D. \Journal{1986}{\MNRAS}{221}{1023}
+\bibitem{} Kolmogorov, A.  \Journal{1941}{\em Compt. Rend. Acad. Sci. USSR}{30}{301}  
+\bibitem{}  Kraichnan, R.H. \Journal{1965}{\em Phys. Fluids}{8}{1385}
+\bibitem{} Kramer, C., Stutzki, J., Rohrig, R., \& Corneliussen, U.
+1998{\AaAp}{329}{249}
+\bibitem{} Langer, W.D., Wilson, R.W., \& Anderson, C.H. 
+\Journal{1993}{\ApJ}{408}{L45}
+\bibitem{} Langer, W. D., Velusamy, T., Kuiper, T.B.H.,
+ Levin, S., \& Olsen, E. \Journal{1995}{\ApJ}{453}{293}
+\bibitem{} Larson, R. \Journal{1992}{\MNRAS}{26}{641} 
+\bibitem{} Lazarian, A. \Journal{1992}{\em Astron. and Astrophys. 
+Transactions}{3}{33}
+\bibitem{} Lazarian, A. \Journal{1994a}{\em Plasma Phys. and Contr. Fusion}
+{36}{1013}
+\bibitem{} Lazarian, A. (1994b), {\em PhD Thesis}, University of Cambridge
+\bibitem{} Lazarian, A. \Journal{1995}{\AaAp}{293}{507} (L95)
+\bibitem{} Lazarian, A. (1999), in 
+{\em Interstellar Turbulence}, eds. Jose Franco \& Alberto Carraminana, CUP, 
+p.95
+\bibitem{} Lazarian, A. \& Pogosyan, D. \Journal{1997}{\ApJ}{491}{200}
+\bibitem{} Lazarian, A. \& Pogosyan, D. \Journal{1999}{\ApJ}{}{in press (LP99),
+astro-ph/9901241}
+\bibitem{} Lazarian, A., \& Vishniac, E. \Journal{1999}{\ApJ}{517}{700}
+\bibitem{} Liszt, H.S., \& Burton, W.B. \Journal{1981}{\ApJ}{243}{778}
+\bibitem{} McKee, C.F. \&  Ostriker, J.P. \Journal{1977}{\ApJ} {218} {148}
+\bibitem{} Melott, A.L., Cohen, A.P., Hamilton, A.J.S., Gott, J.R., Weinberg,
+D.H. \Journal{1989}{\ApJ}{345}{618}
+\bibitem{} Miesch, M.S., \& Bally, J. \Journal{1994}{\ApJ}{429}{645}
+\bibitem{} Montgomery, D. \Journal{1982}{Phys. Scripta}{2}{83}
+\bibitem{} Monin, A.S., \& Yaglom, A.M. (1975), 
+ {\em Statistical Fluid Mechanics: Mechanics of Turbulence}, vol. 2, The MIT Press 
+\bibitem{} Munch, G. \Journal{1958}{\em Rev. Mod. Phys.}{30}{1035}
+\bibitem{} Narayan, R. \Journal{1992}{\em Phil. Trans. Royal Soc.}{341}{151}
+\bibitem{} Narayan, R., \& Goodman, J. \Journal{1989}{\MNRAS}{238}{963}
+\bibitem{} O'Dell, C.R. \Journal{1986}{\ApJ}{304}{767}
+\bibitem{} O'Dell, C.R., \& Castaneda, H.O. \Journal{1987}{\ApJ}{317}{686}
+\bibitem{} Ostriker, E. (1999) in 
+{\em Interstellar Turbulence}, eds. Jose Franco \& Alberto Carraminana, CUP, 
+p.~240
+\bibitem{} Oughton,S., Radler, K.H., Matthaeus,W.H. \Journal{1997}
+{\em Phys. Rev. E}{56}{No.3}, 2875
+\bibitem{} Padoan, P., Rosolowsky, E.W., \& Goodman, A.A. \Journal{1999}{\ApJ}
+{}{submitted}
+\bibitem{} Perault, M., Falgarone, E. \& Puget, J.L. \Journal{ 1986}{\AaAp}
+{157}{139}
+\bibitem{} Pichardo, B., Vazquez-Semadeni, E., Gazol, A., Passot, T.,
+Ballesteros-Paredes, J. \Journal{1999}{\ApJ}{}{submitted}
+\bibitem{} Prunet, S. \& Lazarian, A. (1999), in 
+``Sloan Summit on Microwave Foregrounds'', ed. A. de Oliveira-Costa 
+and M. Tegmark
+\bibitem{} Rauzy, S., Lachieze-Rey, M.,\& Henriksen, R.N. \Journal{1993}
+{\AaAp}{273}{366}
+\bibitem{} Rosolowsky, E.W., Goodman, A.A., Wilner, D.J., \& Williams, J.P.
+\Journal{1999}{\ApJ}{}{in press}
+\bibitem{} Sanders, D.B., Scoville, N.Z., \& Solomon, P.M. \Journal{1985}
+{\ApJ}{289}{373}
+\bibitem{} Scalo, J.M. (1985), in {\em Protostars and Planets II}, 
+eds Black~D.C. and Mathews~M.S.,  Tucson: University of Arizona, 201
+\bibitem{} Scalo, J.M. (1987), in {\em Interstellar Processes}, 
+eds. D.F. Hollenbach \& H.A. Thronson, Reidel, Dordreicht, 349
+\bibitem{} Scalo, J.M., Lazarian, A. \Journal{1996}{\ApJ}{469}{189}
+\bibitem{} Scoccimarro, R. \Journal{1997}{\ApJ}{487}{1}
+\bibitem{} Sethi, S.K., Prunet, S., Bouchet, F.R., 1998, astro-ph/9803158.
+\bibitem{} Shull, J.M. 1987, in {\it Interstellar Processes}, 
+eds Hollenbach~D.J. and Thronson~H.A., 
+ Reidel, Dordrecht, 225
+\bibitem{} Spangler, S.R., \& Gwinn, C.R. \Journal{1990}{\ApJ}{353}{L29}
+\bibitem{} Stanimirovic, S., Staveley-Smith, L., Dickey, J.M.,
+Sault, R.J., \& Snowden, S.L. \Journal{1999}{\MNRAS}{302}{417}
+\bibitem{} Stenholm, L.G. \Journal{1999}{\AaAp}{232}{495}  
+\bibitem{} Stutzki, J., Bensch, F., Heithausen, A., Ossenkopf, V.,
+\& Zielinsky, M. \Journal{1998}{\AaAp}{336}{697}
+\bibitem{} Stutzki, J., \& Gusten, R. \Journal{1990}{\ApJ}{356}{513}
+\bibitem{} Verschuur, G.L. \Journal{1991a}{\em Ap \& SS}{185}{137}
+\bibitem{} Verschuur, G.L. \Journal{1991b}{\em Ap \&SS}{185}{305}
+\bibitem{} Verschuur, G.L. \Journal{1995}{\em Ap\&SS}{227}{187}
+\bibitem{} Vazquez-Semadeni, E. (1999), in  the Proceedings of 
+"The Chaotic Universe", Roma/Pescara, Italy, 1-5 Feb. 1999, eds. 
+V. Gurzadyan and L.
+    Bertone
+\bibitem{} Vazquez-Semadeni, E., \& Passot, T. 
+(1999),
+in {\em Interstellar Turbulence}, eds. Jose Franco \& 
+Alberto Carraminana, CUP, p.~223
+\bibitem{} Wall, W.F., \& Waller, W.H. 1998, in {\em New Horisons 
+from Multi-Wavelength Sky Surveys}, eds. B.J. McLean, D.A. Golombek,
+Hayes, J.J.E. and H.E. Payne, IAU 179,
+Kluwer, p. 191
+\bibitem{} Waller, W.H., Varoxi, F., Boulanger, F., \& Digel, S.W.
+ 1998,  in {\em New Horisons 
+from Multi-Wavelength Sky Surveys}, eds. B.J. McLean, D.A. Golombek,
+Hayes, J.J.E. and H.E. Payne, IAU 179,
+Kluwer, p. 194 
+\bibitem{} Williams, J. (1999) in 
+{\em Interstellar Turbulence}, eds. Jose Franco \& 
+Alberto Carraminana, CUP, p.~190
+\bibitem{} Williams, J., de Gaus, E., \& Blitz, L. 
+\Journal{1994}{\ApJ}{428}{693}
+\bibitem{} Wilson, O.C., Munch, G., Flather, E.M., \& Coffeen, M.F. 
+\Journal{1959} {\em ApJS}{4}{199}
+\bibitem{} Wiserman, J.J. \& Ho, P.T.P. \Journal{1996}{\Nat}{382}{139}
+\bibitem{} Zielinsky, M. \& Stutzki, J. \Journal{1999}{\AaAp}{347}{630}
+\end{thebibliography}
+
+
+\end{document}
+
+\bibitem{Ost:1991}Ostrowski M. \Journal{1991}{\MNRAS}{249}{551}.   
+
+\bibitem{Ost:1996}Ostrowski M. (1996) in Proc. Int. Symp. on
+{\em Extremely High Energy Cosmic Rays}, ed. M.~Nagano, Tanashi
+(p.~377). 
+
+\bibitem{Ost:1990}Ostrowski M. \Journal{1990}{\AAp}{238}{435}.
+
+\bibitem{Abr:1965}Abramowitz M., Stegun I.A., {\em Handbook of Mathematical
+Functions\/} (Dover, New York 1965).
+
+
+
+
+\references
+
+
+\begin{table}[h]
+\begin{center}
+\begin{tabular}{|c|c|c|l|}
+\hline
+AA&BBB &CCCC &DDDd \\
+&$\mit\Delta$ &
+$\mit\Delta-\Gamma$&HHH
+\\ \hline
+\multicolumn{2}{|c|}{CCCC} & & \\
+\cline{1-2}
+$EEE$ &
+$22.711 \times 10^5$ &
+\begin{minipage}{1in}
+$WWWw$ \\ $fff$ 
+\end{minipage} &
+\begin{minipage}{1.5in}
+AAA aaaa BB dfg g g t   gfd g gdf gd gfd gdf gfd gfd g gfd gfdg
+gfdg fd gfd gg fdg fd gg gf hgh gfh e fff fsd ff fd fd fd fd g.
+\end{minipage} \\ 
+& & & \\ \hline
+\end{tabular}
+\end{center}
+\caption{QQQ sdf ff f f df  f fd fd f df ddfh  g gg  ffd ffd df
+ffd fd fd fd  dg dfg fdg fdg fdg fdg fdg tr gtre gre greg  g gr
+g rg rg g g trg tre trg try tyu y e e545 3 fer  dddftr ttr tr tr
+rt  t trgh tr trh rt rh tr h g gg g fg rtg rt rw  gtrg wt tr tr
+tr gtr tr trg rt rett gtre re re ret ret ret ret 54  w254 wtg
+wrt 54 y67u j67 j758 j ry jhg h  d.}
+\label{tab:exp}
+\end{table}
+
+
+\section{Variety of sources and spectral behavior}\label{sec:var}
+
+
+
+\begin{figure}[h]
+\centerline{\psfig{file=fig1.ps,width=7cm,angle=90}}
+\caption{Blazars, $\gamma$-ray burst sources and
+the initial 0.1 sec hard pulse of the March 5, 1979 burst occupy
+distinct regions of the plane defined by source compactness $\ell_c$
+and asymptotic Lorentz factor $\gamma_\infty$.  GRB outflows may contain
+lower-$\gamma_\infty$ ejecta that manifests itself as softer
+tails, precursors and sub-pulses.  
+The Lorentz factors of blazar sources
+are only indirectly constrained by superluminal motion outside the
+$\gamma$-ray emitting region.}
+\label{fig:aaa}
+\end{figure}
+
+
+\section{Final remarks}\label{sec:fin}
+
+
+
+
+
+

arXiv_src_0509_002.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bfee3eb8ad3e2d66f62bbbb75c073f3f7d2c0a0de5e017f1ba9655a11ab8dee2
+size 276111360

arXiv_src_0804_002.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:847f07f8afedb3c9b3acc158f52ad9a64f529766fc3a8151012ace21c3cb3d77
+size 525864960

arXiv_src_0910_004.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d4243b926928c2a7eb9b504845e0c3b12538c149079ec45cafd77532bc972e2
+size 524431360

arXiv_src_1103_005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d892b48dcb727438e31e8b86923ccd6ef5fdcb8223738b5944bb184d1279fc2
+size 558284800

arXiv_src_1107_005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c38240b66231eb3547045433359cc200e3716a8c750bccaf1b38d973dae5d97f
+size 529950720

arXiv_src_1204_009.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:119b3fd77dec4c1b684ef1d080b18928d41a1d81efd17aa91053d8b6d3972e8d
+size 437882880

arXiv_src_1210_006.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68ba601583067ed93f0375129a04039539bb8d1e2d9f18cde4e5c37513e9c621
+size 531159040

arXiv_src_1211_011.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d722598481897b7b13416c2c8d7623d297e470c9ddb2c1e5c287089f0b9a9794
+size 133898240

arXiv_src_1308_007.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:79b042dc03682e196c48cd2ef3f9ceaa4b7ed767ddedea3908c586e90a6e541b
+size 528865280

arXiv_src_1309_010.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39086f0dad437071ee51948d3fc0671a10de1902073067b368aff172dfa9478b
+size 559656960

arXiv_src_1404_011.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28ec5bd6bf2437d59cac22caa2f6720b94911ce4d77623c9be69ec38daa6be6e
+size 547491840

arXiv_src_1409_001.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50336be619bcce361de446b9646a40d4cb1bf27a66fbc07705aba0dfc841abdd
+size 544645120

arXiv_src_1602_003.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae9ab403f17cd673023967efda03d173ef85aa1cc5a60c81c58b55718c6c5446
+size 526510080

arXiv_src_1603_014.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0af4fe4b99b34257eb802a8411d38e4fca593c82c2b1685819bdc31540f3287d
+size 531865600

arXiv_src_1606_003.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94dfff4e66130502695afc88e2eb315e5798eb1eee7a86225534b0b10bff63f2
+size 388935680

arXiv_src_1607_014.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc08332e617a7a2c4058cd8329b52e25d0d4ea85c5fc1f9d94c611f322139bf9
+size 549898240

arXiv_src_1704_023.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ece0661325f346c6ef5047fa3e725e3d475d31569d162ca5f3a17b440e5930c
+size 47278080

arXiv_src_1708_024.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:268e488e942191f8df939b019be311e09cf3a0251602531ac7c9fd0d68fda3ec
+size 495400960

arXiv_src_1712_005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a5227fe276de53a196448e21af54bdc396fbf7db8d31281490d0e6c7d53140b6
+size 535910400

arXiv_src_1802_021.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a20d99c640f547b2d2d1caa2c0e639f6b62c6f1a48681643c584fc55699fd07
+size 539832320

arXiv_src_1804_008.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:676348b8fd1fe119028c9cc522a17ffa3e69eedd67d6ac7945639ef741e8dc31
+size 538419200

arXiv_src_1806_021.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:708289f2ded5006db7a51e8c6a7378d56a728e01387bf34c83af70cbd6d37adf
+size 535470080

arXiv_src_1809_018.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:053d2a934da154cb552e89016d0aae7a6cfe7e9379ad7f3e3ce76a90af5950c7
+size 527452160

arXiv_src_1902_028.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f9bbebd52feb090b201f8be78b9dcfeb033e38a992b94f9af2c717f7056b998
+size 557219840

arXiv_src_1906_028.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24f8a6cdd878f5b263c9ea9cd4ad19032ff2a39caf8c4b78e1ca456309a02264
+size 572108800

arXiv_src_1909_011.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe090a5942159f134455de83af10eee6454e23cc56d514a26d77820618b085c9
+size 571064320

arXiv_src_1911_019.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b05b016892aa69566860c82b4e3fc1a2f18f2b6249cd2cf107a371cedabcf33
+size 545126400

arXiv_src_1912_027.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67b0c824131dadd7e95d341814a24ee2da4fdf4e206c4bffc92e96b5d41607a1
+size 585799680

arXiv_src_2004_005.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5d525c7309061d0b546dd7211ca954977e603f4fe22d9cda604e2dc3ce54f7b9
+size 594073600

arXiv_src_2005_012.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53ed9ecf7bbd652850a01b70f754d379a6096189627a76cb4834d3e8f0d514de
+size 588851200

arXiv_src_2006_050.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dea8858436dc1ea7399bacedb02785f92ab7578e8c0e53f9b294c330ed1ca709
+size 543600640

arXiv_src_2007_047.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f52f375d9c7660fc0d15d00ea794ccb986e06fb1e2961a4f3720703ab5d9ebfc
+size 516003840

arXiv_src_2009_015.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:098fc13ec0b67eeb0324ce69c42104c1fa473f022a83f232787f48044b702216
+size 528240640

arXiv_src_2009_069.tar

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1893d5ae5aec08da87e18379b16185321bb76796701374ecd395c97d25abe872
+size 563691520