7.3 Positive Definite Matrices Definition Very frequently in engineering and statistics applications, we end up dealing with symmetric (& square) matrices which have positive eigenvalues. These matrices are said to be positive definite. Theorem 1 If A is positive definite, then it is invertible and det A > 0. Proof: A (n x n) with eigenvalues 1, 2,…, n det A = 12…n > 0 (since all eigenvalues are pos) Since det A≠ 0, A is invertible. Theorem 2 A symmetric matrix A is positive definite iff XTAX > 0 for every column X ≠ 0 in n. Proof: let PTAP = D = diag(1, 2,…, n) where P-1=PT and the i are the eigenvalues of A (since princ axes thm says that symmetric A is orthogonally diagonalizable). Let Y = PTX = [y1 y2…yn]T and we know A=PDP-1 so XTAX = XT PDP-1 X=YTDY=1y12+…+ nyn2 (*) i>0 since A is pos def. Also, if Y=0, we have the homogeneous system PTX=0 which only has the solution X=0 since P is invertible. But X≠0 so Y≠0. This implies that some yi≠0, so XTAX > 0 (from * above) Theorem 2 (cont) Given XTAX >0 for every X≠0, we will select an X≠0 for which each i≠0: let X = PEj ≠ 0 (where Ej is col j of In-- so we have a col of P). Then Y = PTX = PTPEj=P-1PEj=Ej So recall (*) XTAX = XT PDP-1 X=YTDY=1y12+…+ nyn2 >0 Becomes: j >0 which tells us that A is positive definite. Example If A = CTC where C is invertible, show that A is positive definite. Solution: Try to show that XTAX>0 for all X≠0. XTAX = XTCTCX=(CX)T(CX)=CX•CX=||CX||2>0: reason: with C invertible, the only solution to the homogeneous system CX = 0 is the trivial solution X=0, but X≠0, so CX ≠0, which means ||CX||2>0. Definition Every positive definite matrix A can be factored as A=UTU, where U is upper triangular with positive entries on the main diagonal--called the Cholesky Decomposition (which is a unique factorization). We find the Cholesky Decomposition using elementary row operations (as in the next example). Cholesky Decomposition Algorithm First be sure that your matrix A is positive definite. Then: 1. Add multiples of a row to lower rows in order to carry A to an upper triangular matrix U1. (Only multiply the higher row times a value.) 2. Divide the entries in each row by the square root of the diagonal entry in the row. You will now have U such that A=UTU. It gets messy to verify this in the general case, but we can show it works for various pos. def. matrices. Example Find the Cholesky Decomposition of We only add multiples of a row to rows below it. 2 4 3 0 11 4 0 Can verify that UTU=A 4 3 A 3 5 3 2 11 2 Definition If A is (n x n), and 1r n, let rA be the (r x r) matrix obtained from A by deleting the last (n - r) rows and columns. The matrices 1A, 2A,…,nA = A are called the principle submatrices of A. 1 a11 A a11 , A a21 2 a11 a12 3 a , A 21 a22 a31 a12 a 22 a32 a13 a23 a33 Theorem 3 If A is positive definite, so is rA for each r = 1,2,…,n A P Y r n Proof: A ,letY 0 ; X Q R 0 r 0 < XTAX (since A is pos def) X AX Y T T r A PY T r 0 Y AY Q R0 So YTrAY >0 so rA is positive definite. Lemma Let A be an (n x n) matrix for which det rA>0 for each r=1,2,…,n. Then LA =U where L is (n x n) lower triangular with 1’s on the diagonal, and U is (n x n) upper triangular with positive entries on the diagonal. Proof: (by induction). n=1: A=[a], a>0, so let L=[1] and U=A. n>1: let A=[aij] and note that a11=det1A>0. So A can be taken by row operations that add multiples of row 1 to lower rows to: a11 0 X The row operations to accomplish this are left-mult B by elementary matrix that is lower triangular with 1’s on the diagonal. (verify) Proof (cont) Let L1 be the product of these elementary matrices, which will also be lower triangular with 1’s on the main diagonal. (verify) a11 L1 A 0 X B We could also write L1 and A in block form: r L1 L1 A M 0 r A C r L1 r A P N D E Q R Looking at the two versions of L1A if r2, we can see that r r a11 L1 A (L1 A) 0 r r X a11 B 0 X r1 B Proof (cont) r r a11 L1 A (L1 A) 0 r r X a11 B 0 X r1 B det(rL1rA)= a11det(r-1B) Since det rL1=1 (since L1 is triangular w/ 1’s on diag), det(rL1rA)= det(rL1) det( rA)= detrA So a11det(r-1B)= detrA Since detrA>0 (given) and a11>0 (given), det(r-1B)>0. Proof (cont) By induction, let L0B=U0 be as in the lemma. X a 1 0 1 0 a If L2= 0 L ,then : L L A 0 L 0 11 0 2 1 0 0 11 B X U0 L2L1A is upper triangular with positive diagonal entries since U0 is upper triangular with pos entries on diag (given) and a11>0. Since L2L1 is lower triangular with 1’s on the diagonal, Let L = L2L1 and U = L2L1A which satisfies the condition.• Theorem 4 The following conditions are equivalent for an (n x n) symmetric matrix A: 1. A is positive definite. 2. det rA>0 for each r=1,2,…,n 3. A = UTU where U is upper triangular with positive diagonal entries. Proof: (1) (2): thm 3 says that if A is pos. def, then so is rA for each r=1,2,…,n and thm 1 says that if A is pos def, then det A>0. So done. Proof (cont) (2)(3): Let LA = U0 as in the previous lemma (i.e U0 is upper triangular w/ pos entries on the diagonal, L is lower tri w/ 1’s on diag.) Let D = diag(d1,d2,…,dn) be the diagonal mtx w/ the same diagonal as U0. Then U1=D-1U0 is upper triangular w/ 1’s on the diagonal (we can show this). Then also U0 = DU1 LA = U0 A =L-1U0 = L-1DU1 Also, A = AT = (L-1DU1)T = U1TDT(L-1)T = U1TD(LT)-1 Proof (cont) So L-1DU1 = A = U1TD(LT)-1 So DU1=LU1TD(LT)-1 So DU1LT=LU1TD So DU1LTD-1=LU1T DU1LTD-1 is upper triangular since D is diagonal, U1 is upper triangular, and LT is upper triangular. LU1T is lower triangular w/ 1’s on main diag since both L and U1T are lower tri with 1’s on the main diagonal (verify). Since DU1LTD-1=LU1T they are upper and lower triang (so diag) w/ 1’s on diag (thus we have In) Proof (cont) So LU1T=In L-1 = U1T Let D0= diag( d1 ,..., dn ) So D02 = D. Then U=D0-1U0 is upper triangular with positive diagonal entries since D0-1 is diagonal w/ pos entries and U0 is upper triangular w/ pos entries on diagonal. And UTU =(D0-1U0)T(D0-1U0)=U0T (D0-1)(D0-1)U0 (D0-1)2 is diagonal w/ 1/di on diag which is D-1 So = U0TD-1U0=(D-1U0)TU0 = U1TU0=L-1U0=A (done) Proof (cont) (3)(1) basically the same as in our first example.• Also, proofs of the lemma and theorems in this section prove the legitimacy of the Algorithm for the Cholesky Decomposition.