This preview shows pages 1–4. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full DocumentThis preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: Stochastic Process 11/10/2006 Lecture 7 Minimum Mean Square Error Estimation NCTUEE Summary In this lecture, I will discuss: • Least Squares • Least Squares using SVD • Fundamental Theorem of Estimation • Linear MMSE Notation We will use the following notation rules, unless otherwise noted, to represent symbols during this course. • Boldface upper case letter to represent MATRIX • Boldface lower case letter to represent vector • Superscript ( · ) T and ( · ) H to denote transpose and hermitian (conjugate transpose), respectively • Upper case italic letter to represent RANDOM VARIABLE 71 1 Least Squares Consider the linear model y = H θ + w , where H is a ”known” m × n observation matrix, θ is an n × 1 unknown parameter which may or may not be random, and w is a noise vector. Then, the leastsquares estimator for θ that minimizes the 2norm  y H θ  2 = ( y H θ ) T ( y H θ ) is given by ˆ θ LS = arg min θ  y H θ  2 = ( H T H ) 1 H T y . (1) Remarks: (1) Note that when H is square and nonsingular, the leastsquares esti mator is reduced to ˆ θ LS = H 1 y . (2) The matrix H † = ( H T H ) 1 H T is called the pseudoinverse of H . We have the LS estimator ˆ θ LS = H † y . (3) The matrix H T H must be nonsingular for (1) to hold true, which requires H to be fullrank. In practice, we often solve leastsquares problems using the following system of normal equations: ( H T H ) ˆ θ LS = H T y . (4) Let ˜ y = y H ˆ θ LS . From the normal equations we will find H T ˜ y = . This is known as the orthogonality condition . (5) The minimum leastsquares is found as J min =  y H θ LS  2 = y T ‡ I H ( H T H ) 1 H T · y . 72 2 Geometric Interpretations The leastsquares problem for the linear model y = H θ + w can be interpreted geometrically, from the concept of distance by matrix 2 norm. (1) The received signal y ∈ R m . If the matrix H ∈ R m × n for m ≥ n is fullrank, then the range space S of H is of dimension n , which is a subspace of R m . (2) The LS estimate θ LS is the vector that renders ˆ s = H θ LS the orthogo nal projection of y onto the subspace spanned by the column vectors of H , i.e. the range of H . The orthogonal projection is given by ˆ s = H θ LS = H ( H T H ) 1 H T  {z } , P · y = P · y , where P = H (...
View
Full
Document
This note was uploaded on 11/28/2010 for the course EE 301 taught by Professor Gfung during the Winter '10 term at National Chiao Tung University.
 Winter '10
 GFung

Click to edit the document details