Rank-Deficient Least Squares Problem Βαγγέλης Δούρος EY0619 1 Mini Revision 2 First Remarks (1) z z z z 3 Problem Statement: Given A ∈ R ,m>=n, m r=rank(A)<n and b ∈ R , find x such that min x || Ax − b ||2 Theorem : There are n-r dimensional set of vectors x that minimize || Ax − b ||2 Proof: Let S=Au=0. Dimension of S is n-r. If x minimizes || Ax − b ||2 , so does x+u. Conclusion: The least squares solution is not unique! mxn First Remarks (2) z Theorem: Let σ min = σ min ( A) > 0 ,the smallest singular T A = U Σ V value of . | unT b | i) If x minimizes || Ax − b ||2 then || x ||2 ≥ ,where un σ min is the last column of U ii) changing b to b + δb can change x to x + δx, where || δ x ||2 ≤ || δ b ||2 σ min • Conclusion: When A is nearly rank deficient, solution x is ill-conditioned (and possibly very large) • What can we do? 4 Regularization z z 5 That is to impose extra conditions to the solution so as to convert an ill-conditioned problem to a well-one. Let’s try to approach it to an exactly rankdeficient least squares problem, using the SVD method. SVD: Theorem z Σ1 rxr, U1,V1 with r Columns 6 SVD: Proof (1) 7 SVD: Proof (2) z 8 Key point: The minimum norm solution x is unique and may be well-conditioned provided σmin is not too small Problems… z z z z 9 Let A = ⎛ ⎜ 1 0 ⎞, ⎟ 0 0 ⎝ ⎠ b = [1,1]T . Applying the previous theorem we take the unique solution x = [1, 0] and the condition number is 1. 1 0 ⎞ b = [1,1]T Let A = ⎛⎜ . Applying the previous ⎟, ⎝0 ε ⎠ T x = [1,1/ ε ] theorem we take the unique solution and the condition number is 1/ε >> 1. So, for a slight different example things are much more difficult… It needs decisiveness!!! T Truncated SVD z z We are going to define a measure tol, setting each singular value smaller than tol to zero!! Let U Σ V T the truncated SVD. As || U Σ V Τ − U ΣV Τ ||2 =|| U (Σ − Σ)V Τ ||2 ≤|| U ||2 || (Σ − Σ) ||2 || V ||2 =|| (Σ − Σ) ||2 ≤ tol • 10 we can use the truncated method instead of the original. The advantage: Not only do we compute the minimum solution, but also we minimize the condition number! QR Method ⎛ R11 A = QR = Q ⎜ ⎝ 0 R12 ⎞ ⎟ and R22 ⎠ R11 is of order r z Let z Provided that || R22 ||2 0 and [Q, Q ] square and orthogonal T T ⎡ ⎤ ⎡ Q Rx − Q b⎤ 2 2 2 || Ax − b ||2 =|| ⎢ T ⎥ ( Ax − b) ||2 =|| ⎢ T ⎥ ||2 = ⎣⎢Q ⎦⎥ ⎣⎢ −Q b ⎦⎥ =|| Rx − QT b ||2 + || Q T b ||2 2 2 Q = [Q1 , Q2 ], x = [ x1 , x2 ]T T 2 || Ax − b || =|| R11 x1 + R12 x2 − Q1 b || + || Q b || + || Q b ||2 2 2 z 11 Is minimized choosing T 2 2 T 2 2 2 ⎡ R11−1 (Q1T b − R12 x2 ) ⎤ x=⎢ ⎥ ⎣ x2 ⎦ QR Method: Bad Scenario z Let z Though condition number of the matrix increases rapidly, || R22 ||2 is not small!! (try it with matlab) So original QR failures to recognize nearly rank deficient R! • 12 ⎛1/ 2 1 0 ⎞ ⎜ ⎟ A=⎜ % 1 ⎟ ⎜ ⎟ 1/ 2 0 ⎝ ⎠ A=QR, Q=I,R=A . QR with Pivoting z z z z 13 We factorize AP=QR, P: permutation matrix At step i, we select from the unfinished part of A (columns i Æn, rows i Æm) the column of largest norm and we exchange it with the i-th column. So, we try to keep R11 as well-conditioned as possible and R22 as small as possible However, this greedy strategy fails to some types of matrices (Kahan matrices) RRQR z Let A mxn and singular values σ 1 ≥ σ 2 ≥ ... ≥ σ n ≥ 0 R ⎞ ⎛R AP = QR = Q ⎜ ⎟ , R 11 is of order r ⎝ 0 R ⎠ RRQR=rank-revealing QR factorization is the σ cond ( R ) = ,|| R || = σ ( R ) = σ one where σ Find a permutation that maximizes σ m in ( R1 1 ) Find a permutation that minimizes σ max ( R22 ) 11 • 12 22 1 11 • • 14 22 r 2 max 22 r +1 Which Method Should I Choose? z z z z z 15 Tradeoff between reliability and speed QR without pivoting is fastest but less reliable… … SVD is slowest but most reliable… QR with pivoting/RRQR are in the middle. Generally speaking, if m>>n, all algorithms cost about the same References z z z z 16 James W. Demmel, “Applied Numerical Linear Algebra” http://www.ms.uky.edu/~bai/Class/lecture19. ps http://www.cse.uiuc.edu/heath/scicomp/notes /chap03.pdf http://www.mcs.anl.gov/prism/lib/techsrc/wn3 0.ps Ευχαριστώ! 17
© Copyright 2026 Paperzz