Transpose and Adjoint Lecture 2 Orthogonal Vectors and Matrices, Norms MIT 18.335J / 6.337J Introduction to Numerical Methods Per-Olof Persson (persson@mit.edu) September 10, 2007 • For real A, the transpose of A is obtained by interchanging rows/columns a11 a12 a a a 11 21 31 T A= a21 a22 =⇒ A = a12 a22 a32 a31 a32 • The adjoint or hermitian conjugate also takes complex conjugate a11 a12 a a a 11 21 31 ∗ A= a21 a22 =⇒ A = a12 a22 a32 a31 a32 • If real A = AT , then A is symmetric. If A = A∗ , then A is hermitian. 1 2 Inner Product In MATLAB • Inner product of two column vectors x, y ∈ Cm x∗ y = m X Quantity x i yi kxk = x∗ x = i=1 • Angle α between x, y cos α = A.’ Transpose only Adjoint A A’ Transpose + complex conjugate Inner product x∗ y x’*y dot(x,y) Length kxk sqrt(x’*x) or norm(x) ∗ • Euclidean length of x m X Comment Transpose A i=1 √ MATLAB Syntax T 2 |xi | !1/2 or ’* assumes column vectors ’* assumes column vector x∗ y kxkkyk 3 4 Orthogonal Vectors Orthogonal and Unitary Matrices • The vectors x, y ∈ Rm are orthogonal if • A square matrix Q ∈ Cm×m is unitary (orthogonal in real case), if x∗ y = 0 Q∗ = Q−1 • For unitary Q • The sets of vectors X, Y are orthogonal if every x ∈ X is orthogonal to every y ∈ Y • Interpretation of unitary-times-vector product: • A set of (nonzero) vectors S is orthogonal if vectors pairwise orthogonal, i.e., for x, y Q∗ Q = I , or qi∗ qj = δij ∈ S, x 6= y ⇒ x∗ y = 0 and orthonormal if, in addition, x = Q∗ b = solution to Qx = b = the vector of coefficients of the expansion of b in the basis of columns of Q every x ∈ S has kxk = 1 5 6 Vector Norms Preservation of Geometry Structure • A norm is a function k · k : Cm → R satisfying • Inner product is preserved under multiplication by unitary Q (Qx)∗ (Qy) = x∗ Q∗ Qy = x∗ y (1) kxk ≥ 0, and kxk = 0 only if x = 0 (2) kx + yk ≤ kxk + kyk • Therefore, lengths of vectors and angles between vectors are preserved • A real orthogonal Q is either a rigid rotation or reflection Rotation • Example: The Euclidean length function √ kxk2 = x∗ x Reflection Qu Qv v Qv v Qu u (3) kαxk = |α| kxk u • k · k2 is a special case of the p-norms !1/p m X p kxkp = |xi | (1 ≤ p < ∞) i=1 7 8 Examples of Vector Norms Verification of Norm Conditions • Show that kxk1 = kxk1 = kxk2 = m X i=1 (1) |xi | m X |xi |2 i=1 !1/2 i=1 |xi | is a norm kxk1 = |x1 | + |x2 | + · · · + |xm | ≥ 0, equality only if x = 0 = √ x∗ x (2) kx + yk1 = kxk∞ = max |xi | 1≤i≤m kxkW = kW xk2 = Pm m X i=1 |wi xi |2 !1/2 = m X i=1 m X i=1 |xi + yi | ≤ |xi | + m X i=1 m X i=1 |xi | + |yi | |yi | = kxk1 + kyk1 (3) m X kαxk1 = i=1 |αxi | = m X i=1 |α| |xi | = |α| m X i=1 |xi | = |α| kxk1 9 10 Induced Matrix Norms Examples of Matrix Norms • For a matrix A ∈ Cm×n , the induced matrix norm is kAk(m,n) = sup x∈Cn x6=0 kAxk(m) = kxk(n) sup kAxk(m) x∈Cn kxk(n) =1 where k · k(n) and k · k(m) are given vector norms kAk1 = max 1≤j≤n kAk∞ = max 1≤i≤m • The “maximum stretching” by A kAkF = kAk2 = 11 m X i=1 n X j=1 m X n X i=1 j=1 p |aij | “maximum column sum” |aij | “maximum row sum” 2 |aij | !1/2 λmax (A∗ A) The Frobenius norm More later 12 In MATLAB Properties of Matrix Norms • Bound on Matrix Product - Induced norms and Frobenius norm satisfy kABk ≤ kAkkBk but some matrix norms do not! • Invariance under Unitary Multiplication - For A ∈ Cm×n and unitary Q ∈ Cm×m , we have kQAk2 = kAk2 , Proof. Since kQxk2 kQAkF = kAkF = kxk2 (inner product is preserved), the first result follows from the definition of induced norm. For the Frobenius norm, p tr((QA)∗ (QA)) = p = tr(A∗ A) = kAkF kQAkF = 13 p tr(A∗ Q∗ QA) Quantity MATLAB Syntax kxk1 sum(abs(x)) or norm(x,1) kxk2 sqrt(x’*x) or norm(x) kxkp sum(abs(x).ˆ p).ˆ (1/p) or norm(x,p) kxk∞ max(abs(x)) or norm(x,inf) kAk1 max(sum(abs(A),1)) or norm(A,1) kAk2 norm(A) kAk∞ max(sum(abs(A),2)) or norm(A,inf) kAkF sqrt(A(:)’*A(:)) or norm(A,’fro’) 14