III. Multi-Dimensional Random Variables and Application in Vector Quantization Matrix Notation for Representing Vectors Assume X is a two-dimensional vector, then in matrix notation it is represented as: X x1 x 2 The norm of any vector X || X || could be computed using the dot product as follows: 2 x2 X UX x1 X = XX T For any vector X the unit vector UX is defined as X UX = X © Tallal Elshabrawy 2 Projection and Dot Product We would like to evaluate the projection of vector Y over vector X using the matrix notation (i.e, we would like to compute the value m), m x1 y1 x 2 y 2 x12 x 22 XY T m X For proof see lecture notes Y=[y1 y2] X=[x1 x2] a x12 x 22 m m © Tallal Elshabrawy 3 Principle Component Analysis Suppose we have a number of samples in a two dimensional space and we would like to identify a unit vector U that crosses the origin for which these samples are the closest. This could be achieved by finding the vector for which the sum of squares of distances to such vector is minimized (i.e., find U that minimizes d12+d22+d32+d42+…) U X4=[x14 x24] X1=[x11 x21] d4 d3 d1 d2 X3=[x13 x23] X2=[x12 x22] © Tallal Elshabrawy 4 Principle Component Analysis From the Pythagoras theorem, we could argue that the closest vector to the observed samples is also equivalently a problem of finding the vector over which the sum of projection of the sample points square on it is maximized. l min d d d d ... 2 1 2 2 2 3 2 4 d is equivalent to m max m12 m 22 m32 m 24 ... max X1 U X U X U X U T 2 © Tallal Elshabrawy T 2 2 T 2 3 T 2 4 ... l2 d 2 m 2 For constant l d m 5 Principle Component Analysis Define X1 x11 X x 2 12 X x X 3 13 X 4 x14 . . . . x 21 x 22 x 23 x 24 . . X4=[x14 x24] X1=[x11 x21] U X3=[x13 x23] X2=[x12 x22] x12 x13 x14 . . x u 2 11 x x x x . . 22 23 24 21 UX T u1 x11 u 2 x 21 u1 x12 u 2 x 22 u1x13 u 2 x 23 UX T u1 u1x14 u 2 x 24 . . 2 2 2 2 UX T XU T u1 x11 u 2 x 21 u1 x12 u 2 x 22 u1 x13 u 2 x 23 u1x14 u 2 x 24 ... UX T XU T X1 U T X 2 U T X 3 U T X 4 U T 2 © Tallal Elshabrawy 2 2 2 6 Principle Component Analysis Therefore the unit vector U (i.e., UUT=1) that is closest to sample points X1, X2, X3, X4 satisfies max UX XU T T Suppose the maximum value is λ X4=[x14 x24] X1=[x11 x21] U X3=[x13 x23] X2=[x12 x22] UX T XUT λ λUUT U λU T U X T XUT λUT 0 X T X U T λUT The equation above is an eigen vector problem for the matrix XTX which means that the maximum we are seeking λ must solve the equation above and it is therefore one of the eigen values (maximum eigen value) for the SQUARE matrix XTX © Tallal Elshabrawy 7 Two-Dimensional Random Variables Assume X is a two-dimensional random variable, then in matrix notation X=[X1 X2] What does it mean that X is a two-dimensional random variable? It means that there is this experiment/phenomenon that could be expressed in terms of 2 random variables X1 and X2 Example: Weather (W) could be categorized as a twodimensional random variable if we characterize it by Temperature (T) and Humidity (H). (i.e., W = [T, H]) © Tallal Elshabrawy 8 Parameters of Two-Dimensional R.V. Mean of any of the Random Variables E Xi k Pr Xi k if Xi depict discrete random variables i=1,2 Xi E Xi yf y dy Xi if Xi depict continuous random variables i=1,2 Xi Variance of any of the Random Variables 2 2 var X i E Xi E Xi k E X i Pr X i k Xi i=1,2 if Xi depict discrete random variables 2 2 var X i E X i E X i y E X i f Xi y dy X i=1,2 i if Xi depict continuous random variables © Tallal Elshabrawy 9 Parameters for Relation Between R.V.s Correlation Corr X1 , X 2 E X1X 2 Covariance Cov X1 , X 2 E X1 E X1 X 2 E X 2 © Tallal Elshabrawy 10 Covariance Matrix cov X1 , X 2 var X1 RX var X 2 cov X 2 , X1 E X' 2 E X' X' 1 2 1 RX 2 E X1' X '2 E X '2 where X1' X1 E X1 and X'2 X2 E X2 © Tallal Elshabrawy 11 2-D R.V. & Principle Component Analysis Assume X is a two-dimensional random variable, then in matrix notation X=[X1 X2] Assume X’ is a two-dimensional random variable, then in matrix notation X’=[X’1 X’2] where X1’=X1 - E[X’1], X2’=X2 - E[X’2] The vector U that is closest to sample points of the twodimensional random variable X is the eigen vector that corresponds to the maximum eigen value for the Covariance Matrix Rx’. i.e., U solves R X ' U T λU T © Tallal Elshabrawy (Proof in lecture notes) 12 © Tallal Elshabrawy 13 © Tallal Elshabrawy 14 © Tallal Elshabrawy 15 © Tallal Elshabrawy 16 © Tallal Elshabrawy 17
© Copyright 2026 Paperzz