% I've made the source of this available September 2004 as I realised people % might want to adapt this for their own courses. I guess it's LaTeX2E. I did % some not very pretty fudging to fit stuff in how I wanted it. I used psnup to % fit the four pages onto a double sided sheet of A4 paper. Not pretty, but was % good enough. % % I'd be interested to see any improved or alternative versions of this % document. Also LaTeX wizards: please feel free to show me how to do this % better. % % Iain Murray, http://www.gatsby.ucl.ac.uk/~iam23/ \documentclass[a4paper]{article} \usepackage{amsmath,amsthm,amsfonts} \usepackage{graphicx} \usepackage[bottom]{footmisc} % normal footnotes weren't behaving % Text position and margin fudging \addtolength{\oddsidemargin}{-0.7in} \addtolength{\evensidemargin}{-0.7in} \addtolength{\marginparwidth}{+0.7in} \addtolength{\marginparpush}{+5pt} \let\oldmarginpar\marginpar \renewcommand\marginpar[1]{\-\oldmarginpar[\raggedleft\small\sf #1]{\raggedright\small\sf #1}} % I use \prob for continuous distributions and \Prob for discreet (except when % I forget). I can then change these to reformat all probabilities to something % fancier if I want. \newcommand{\prob}[1]{p\left( #1 \right)} \newcommand{\Prob}[1]{P\left( #1 \right)} % Common stuff for linear algebra \newcommand{\eye}{\mathbb I} \newcommand{\bx}{{\bf x}} \newcommand{\by}{{\bf y}} \newcommand{\be}{{\bf e}} %calculus \newcommand{\ddd}[2]{\frac{{\rm d}^{2} #1}{{\rm d} #2 ^{2}}} \newcommand{\dd}[2]{\frac{{\rm d} #1}{{\rm d} #2}} % Misc \newcommand{\ith}{^{\left( i \right)}} \newcommand{\te}{\!=\!} % thin equals \title{Background material crib-sheet} \author{Iain Murray $<${\tt i.murray+ta@gatsby.ucl.ac.uk}$>$, October 2003} \date{} \begin{document} \maketitle {\small \em Here are a summary of results with which you should be familiar. If anything here is unclear you should to do some further reading and exercises.} \section{Probability Theory} {\small \em Chapter 2, sections 2.1--2.3 of David MacKay's book covers this material:\\ \indent {\tt http://www.inference.phy.cam.ac.uk/mackay/itila/book.html}} \bigskip \noindent% \marginpar{Notation}\ The probability a discrete variable $A$ takes value $a$ is:\quad $0 \leq \Prob{A\!=\!a} \leq 1$ \bigskip \noindent \marginpar{Alternatives}\ Probabilities of alternatives add: $\Prob{A \!=\! a~{\mathrm or}~a^\prime}=\Prob{A \!=\! a} + \Prob{A \!=\! a^\prime}$ \bigskip \noindent \marginpar{Normalisation}\ The probabilities of all outcomes must sum to one: $\displaystyle \sum_{\mathrm{all~possible~} a} \Prob{A\!=\!a} = 1$ \bigskip \noindent \marginpar{Joint Probability}\ $\Prob{A\!=\!a,B\!=\!b}$ is the joint probability that both $A\!=\!a$ and $B\!=\!b$ occur. \bigskip \noindent \marginpar{Marginalisation}\ Variables can be ``summed out'' of joint distributions: \[ \Prob{A\te a} = \sum_{\mathrm{all~possible~} b} \Prob{A\!=\!a,B\!=\!b} \] \bigskip \noindent \marginpar{Conditional Probability}\ $\Prob{A\!=\!a|B\!=\!b}$ is the probability $A\!=\!a$ occurs given the knowledge $B\!=\!b$. \bigskip \noindent \marginpar{Product Rule}\ $\Prob{A\!=\!a,B\!=\!b} = \Prob{A\!=\!a}\Prob{B\!=\!b|A\!=\!a} = \Prob{B\!=\!b}\Prob{A\!=\!a|B\!=\!b}$ \bigskip \noindent \marginpar{Independence}\ The following hold, for all $a$ and $b$, {\em \bf if and only if $A$ and $B$ are independent}:\\ \[ \begin{array}{rcl} \Prob{A\!=\!a|B\!=\!b}&=&\Prob{A\!=\!a} \\ \Prob{B\!=\!b|A\!=\!a}&=&\Prob{B\!=\!b} \\ \Prob{A\!=\!a,B\!=\!b}&=&\Prob{A\!=\!a}\Prob{B\!=\!b}. \end{array} \] Otherwise the product rule above {\em must} be used. \bigskip \noindent \marginpar{Bayes Rule}Bayes rule can be derived from the above: \[ \Prob{A\te a|B\te b,{\mathcal H}} = \frac{\Prob{B\te b|A\te a,{\mathcal H}}\Prob{A\te a|{\mathcal H}}}{\Prob{B\te b|{\mathcal H}}} \propto \Prob{A\te a,B\te b|{\mathcal H}} \] Note that here, as with any expression, we are free to condition the whole thing on any set of assumptions, ${\mathcal H}$, we like. %Summing over all possible $a$ will give the ``evidence'' $\Prob{B\te b|{\mathcal H}}$, Note $\sum_a\Prob{A\te a,B\te b|{\mathcal H}}=\Prob{B\te b|{\mathcal H}}$ gives the normalising constant of proportionality. \pagebreak \noindent \marginpar{Continuous variables}All the above theory basically still applies to continuous variables if sums are converted into integrals\footnote{Integrals are the equivalent of sums for continuous variables. Eg: $\sum_{i=1}^n f(x_i)\Delta x$ becomes the integral $\int_a^b f(x) \mathrm{d}x$ in the limit $\Delta x \rightarrow 0$, $n\rightarrow \infty$, where $\Delta x = \frac{b-a}{n}$ and $x_i = a + i\Delta x $. Find an A-level text book with some diagrams if you have not seen this before. }. The probability that $X$ lies between $x$ and $x\!+\!\mathrm{d}x$ is $\prob{x}\mathrm{d}x$, % yeah I know x+ dx isn't very rigorous. pah. where $\prob{x}$ is a {\em probability density function} with range $[0,\infty ]$. \marginpar{\vspace{12pt}Continuous versions of some results} \[ \Prob{x_1\!<\!X\!<\!x_2} = \int_{x_1}^{x_2} \prob{x} \mathrm{d}x \,, \;\; \int_{-\infty}^{\infty} \prob{x} \mathrm{d}x = 1 \;\;\mbox{and}\;\; \prob{x} = \int_{-\infty}^{\infty} \prob{x,y} \mathrm{d}y . \] \noindent \marginpar{Expectations}The expectation or mean under a probability distribution is: \[ \left< f(a) \right> = \sum_a \Prob{A\te a} f(a) \;\;\;\mbox{or}\;\;\; \left< f(x) \right> = \int_{-\infty}^{\infty} \prob{x} f(x) \mathrm{d}x \] \section{Linear Algebra} {\em \small This is designed as a prequel to Sam Roweis's ``matrix identities'' sheet:\\ \indent {\tt http://www.cs.toronto.edu/\~{}roweis/notes/matrixid.pdf}} \bigskip \noindent Scalars are individual numbers, vectors are columns of numbers, matrices are rectangular grids of numbers, eg: \[ x=3.4, \;\;\;\;\;\; \bx = \left( \begin{array}{c}x_1 \\ x_2 \\ \vdots \\ x_n \end{array} \right), \;\;\;\;\;\; A = \left( \begin{array}{cccc} A_{11} & A_{12} & \cdots & A_{1n} \\ A_{21} & A_{22} & \cdots & A_{2n} \\ \vdots & \vdots & \ddots & \vdots \\ A_{m1} & A_{m2} & \cdots & A_{mn} \\ \end{array} \right) \] \bigskip \noindent \marginpar{Dimensions}In the above example $x$ is $1\times 1$, $\bx$ is $n\times 1$ and $A$ is $m\times n$. \bigskip \noindent \marginpar{Transpose}The transpose operator, ${}^\top$ ( ${}^\prime$ in Matlab), swaps the rows and columns: \[ x^\top = x, \;\;\;\;\;\; \bx^\top = \left( \begin{array}{cccc} x_1 & x_2 & \cdots & x_n \end{array} \right), \;\;\;\;\;\; {\left(A^\top\right)}_{ij} = A_{ji} \] \bigskip \noindent \marginpar{Multiplication}Quantities whose inner dimensions match may be ``multiplied'' by summing over this index. The outer dimensions give the dimensions of the answer. \[ A\bx \mbox{~has elements~} {(A\bx)}_i = \sum_{j=1}^n A_{ij} \bx_j \;\;\; \mbox{and~} \;\;\; {(AA^\top)}_{ij} = \sum_{k=1}^n A_{ik} {\left(A^\top\right)}_{kj} = \sum_{k=1}^n A_{ik} A_{jk} \] \marginpar{Check Dimensions}All the following are allowed (the dimensions of the answer are also shown): \[ \begin{array}{c} \bx^\top \bx \\ \mbox{\small $1\times 1$} \\\mbox{\small scalar} \end{array} \;\;\;\;\;\; \begin{array}{c} \bx \bx^\top \\ \mbox{\small $n\times n$} \\\mbox{\small matrix} \end{array} \;\;\;\;\;\; \begin{array}{c} A \bx \\ \mbox{\small $m\times 1$} \\\mbox{\small vector} \end{array} \;\;\;\;\;\; \begin{array}{c} A A^\top \\ \mbox{\small $m\times m$} \\\mbox{\small matrix} \end{array} \;\;\;\;\;\; \begin{array}{c} A^\top A \\ \mbox{\small $n\times n$} \\\mbox{\small matrix} \end{array} \;\;\;\;\;\; \begin{array}{c} \bx^\top A \bx \\ \mbox{\small $1\times 1$} \\\mbox{\small scalar} \end{array} \;\;\;\;\;\; , \] while $\bx \bx$, $AA$ and $\bx A$ {\em do not make sense} for $m\neq n\neq 1$. Can you see why? %{\em \bf Checking dimensions keeps you sane}. \bigskip \noindent \marginpar{Multiplication by scalar}An exception to the above rule is that we may write: $xA$. Every element of the matrix $A$ is multiplied by the scalar $x$. \bigskip \noindent \marginpar{Easily proved results}Simple and valid manipulations: \[ (AB)C= A(BC) \;\;\;\;\;\; A(B+C)= AB + AC \;\;\;\;\;\; (A+B)^\top = A^\top + B^\top \;\;\;\;\;\; (AB)^\top = B^\top A^\top \] Note that $AB\neq BA$ in general. \subsection{Square Matrices} Now consider the square $n\times n$ matrix $B$. \bigskip \noindent \marginpar{Diagonal matrices, the Identity}All off-diagonal elements of diagonal matrices are zero. The ``Identity matrix'', which leaves vectors and matrices unchanged on multiplication, is diagonal with each non-zero element equal to one. \[ \begin{array}{c} \begin{array}{rcl} B_{ij} = 0 \mbox{~if~} i \! \neq\! j &\Leftrightarrow& \mbox{``$B$ is diagonal''} \\ \eye_{ij} = 0 \mbox{~if~} i \! \neq\! j \mbox{~and~} \eye_{ii}=1\;\;\forall i &\Leftrightarrow& \mbox{``$\eye$ is the identity matrix''} \end{array} \\ \eye \bx = \bx \;\;\;\;\;\; \eye B = B = B \eye \;\;\;\;\;\; \bx^\top \eye = \bx^\top \end{array} \] \noindent \marginpar{Inverses}Some square matrices have inverses: \[ B^{-1}B= BB^{-1}=\eye \;\;\;\;\;\; {\left(B^{-1}\right)}^{-1}= B \, , \] which have these properties: \[ (BC)^{-1} = C^{-1}B^{-1} \;\;\;\;\;\; {\left(B^{-1}\right)}^\top = {\left(B^\top\right)}^{-1} \] \noindent \marginpar{Solving Linear equations}Linear simultaneous equations could be solved (inefficiently) this way: \[ \mbox{if~} B\bx = \by \mbox{~then~} \bx = B^{-1}\by \] \noindent Some other commonly used matrix definitions include: \marginpar{\vspace{12pt}Symmetry} \[ B_{ij} = B_{ji} \Leftrightarrow \mbox{``$B$ is symmetric''} \] \vspace{-12pt} \marginpar{\vspace{12pt}Trace} \[ \mathrm{Trace}(B)= \mathrm{Tr}(B) = \sum_{i=1}^n B_{ii}=\mbox{``sum of diagonal elements''} \] \marginpar{A Trace Trick}Cyclic permutations are allowed inside trace. Trace of a scalar is a scalar: \[ \mathrm{Tr}(BCD)= \mathrm{Tr}(DBC) = \mathrm{Tr}(CDB) \;\;\;\;\;\; \bx^\top B\bx = \mathrm{Tr}(\bx^\top B\bx) = \mathrm{Tr}(\bx \bx^\top B) \] \bigskip \noindent \marginpar{Determinants}The determinant\footnote{This section is only intended to give you a flavour so you understand other references and Sam's crib sheet. More detailed history and overview is here: {\tt http://www.wikipedia.org/wiki/Determinant}} is written Det$(B)$ or $|B|$. It is a scalar regardless of $n$. \[ |BC|=|B||C| \, , \;\;\;\;\;\; |x|= x \, , \;\;\;\;\;\; |xB| %= |x\eye||B| = x^n |B| \, , \;\;\;\;\;\; \left|B^{-1}\right| = \frac{1}{|B|} \, . \] It {\em determines} if $B$ can be inverted: $|B|\!=\!0 \Rightarrow B^{-1}$ undefined. If the vector to every point of a shape is pre-multiplied by $B$ then the shape's area or volume increases by a factor of $|B|$. It also appears in the normalising constant of a Gaussian. For a diagonal matrix the volume scaling factor is simply the product of the diagonal elements. In general the determinant is the product of the eigenvalues. \marginpar{\vspace{12pt}Eigenvalues, Eigenvectors} \[ B \be\ith = \lambda\ith \be\ith \Leftrightarrow \mbox{``$\lambda\ith$ is an eigenvalue of $B$ with eigenvector $\be\ith$''} \] \[ |B| = \prod \mbox{eigenvalues} \;\;\;\;\;\; \mbox{Trace}(B) = \sum \mbox{eigenvalues} \] If $B$ is real and symmetric (eg a covariance matrix) the eigenvectors are orthogonal (perpendicular) and so form a basis (can be used as axes). \section{Differentiation} {\small \em Any good A-level maths text book should cover this material and have plenty of exercises. Undergraduate text books might cover it quickly in less than a chapter.} \bigskip \noindent \marginpar{Gradient}The gradient of a straight line $y\!=\!mx\!+\!c$ is a constant $y^\prime = \frac{y(x+\Delta x)-y(x)}{\Delta x}=m$. %for any $x$ or $\Delta x$. \bigskip \noindent \marginpar{Differentiation}Many functions look like straight lines over a small enough range. The gradient of this line, the derivative, is not constant, but a new function: \[ y^\prime(x) = \dd{y}{x} = \lim_{\Delta x\to 0} \frac{y(x\!+\!\Delta x)-y(x)}{\Delta x}\;, \;\;\;\;\; \parbox{3.2cm}{which could be\\differentiated again:}\;\; y^{\prime\prime} = \ddd{y}{x} = \dd{y^\prime}{x} \] \bigskip \noindent \marginpar{Standard derivatives}The following results are well known ($c$ is a constant): \[ \begin{array}{c}f(x):\\f^\prime(x):\end{array} \;\;\;\;\;\; \begin{array}{c}c\\0\end{array} \;\;\;\;\;\; \begin{array}{c}cx\\c\end{array} \;\;\;\;\;\; \begin{array}{c}cx^n\\cnx^{n-1}\end{array} \;\;\;\;\;\; \begin{array}{c}\log_e(x)\\1/x\end{array} \;\;\;\;\;\; \begin{array}{c}\exp(x)\\\exp(x)\end{array} % They won't use these: %\;\;\;\;\;\; %\begin{array}{c}\sin(x)\\\cos(x)\end{array} %\;\;\;\;\;\; %\begin{array}{c}\cos(x)\\-\sin(x)\end{array} %\;\;\;\;\;\; \;\;\; . \] \bigskip \noindent \marginpar{Optimisation}At a maximum or minimum the function is rising on one side and falling on the other. In between the gradient must be zero. Therefore \[ \mbox{maxima and minima satisfy:~} \dd{f(x)}{x}=0 \;\;\;\;\;\; \mbox{or} \;\;\;\;\;\; \dd{f(\bx)}{\bx}={\mathbf 0} \; \Leftrightarrow \; \dd{f(\bx)}{x_i}=0\;\;\;\forall i \] If we can't solve this we can evolve our variable $x$, or variables $\bx$, on a computer using gradient information until we find a place where the gradient is zero. \bigskip \noindent \marginpar{Approximation}A function may be approximated by a straight line\footnote{More accurate approximations can be made. Look up Taylor series.} about any point $a$. \[ f(a+x) \approx f(a) + x f^\prime(a)\;, \;\;\;\;\;\; \mbox{~eg:~} %\;\;\;\;\;\; \log(1+x) \approx \log(1+0) + x\frac{1}{1+0} = x \] \bigskip \noindent \marginpar{Linearity}The derivative operator is linear: \[ \dd{(f(x)+g(x))}{x} = \dd{f(x)}{x} + \dd{g(x)}{x}\;, \;\;\;\;\;\; \mbox{~eg:~}\; \dd{\left(x+\exp(x)\right)}{x} = 1 + \exp(x). \] \bigskip \noindent \marginpar{Product Rule}Dealing with products is slightly more involved: \[ \dd{\left(u(x)v(x)\right)}{x} = v\dd{u}{x} + u\dd{v}{x}\;, \;\;\;\;\; \mbox{~eg:~}\; \dd{\left(x\cdot\exp(x)\right)}{x} = \exp(x) + x\exp(x). \] \bigskip \noindent \marginpar{Chain Rule}The ``chain rule'' $ \displaystyle \dd{f(u)}{x} = \dd{u}{x}\dd{f(u)}{u}, $ allows results to be combined. \[ \begin{split} \mbox{For example:~} \dd{\exp\left(ay^m\right)}{y} &= \dd{\left(ay^m\right)}{y} \cdot \dd{\exp\left(ay^m\right)}{\left(ay^m\right)} \;\;\; \mbox{``with~$u=ay^m$''}\\ &= amy^{m-1} \cdot \exp\left(ay^m\right) \end{split} \] \marginpar{\em Exercise}If you can't show the following you could do with some practice: \[ \dd{}{z} \left[ \frac{1}{(b+cz)} \exp(az) + e \right] = \exp(az) \left(\frac{a}{b+cz}-\frac{c}{(b+cz)^2}\right) \] Note that $a,b,c$ and $e$ are constants, that $\frac{1}{u}=u^{-1}$ and this is hard if you haven't done differentiation (for a long time). Again, get a text book. \end{document}