Orthogonal Complements c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 1 / 22 The orthogonal complement of a vector space S ⊆ Rn , denoted S ⊥ , is S ⊥ = {x ∈ Rn : x0 y = 0 ∀ y ∈ S}. Is S ⊥ is a vector space? c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 2 / 22 ∀ c1 , c2 ∈ R, x1 , x2 ∈ S ⊥ , (c1 x1 + c2 x2 )0 y = c1 x10 y + c2 x02 y = 0 whenever y ∈ S. Thus x1 , x2 ∈ S ⊥ , c1 , c2 ∈ R =⇒ c1 x1 + c2 x2 ∈ S ⊥ . It follows that S ⊥ is a vector space. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 3 / 22 Suppose a vector space S is a subset of Rn . Show the following: (a) S ∩ S ⊥ = {0} (b) dim(S) + dim(S ⊥ ) = n. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 4 / 22 Proof of (a) Suppose x ∈ S ∩ S ⊥ . Then x0 x = 0, which implies that n X xi2 = 0 =⇒ xi = 0 ∀ i = 1, . . . , n i=1 =⇒ x = 0. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 5 / 22 Proof of (b) Let r = dim(S) and let a1 , . . . , ar be a basis for S. If we define A = [a1 , . . . , ar ], then C(A) = S. Now note that S ⊥ = N (A0 ) as follows: c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 6 / 22 x ∈ S ⊥ =⇒ x0 ai = 0 ∀ i = 1, . . . , r =⇒ x0 [a1 , . . . , ar ] = 00 =⇒ x0 A = 00 =⇒ A0 x = 0 =⇒ x ∈ N (A0 ) ∴ S ⊥ ⊆ N (A0 ). Conversely, x ∈ N (A0 ) =⇒ A0 x = 0 =⇒ x0 A = 00 =⇒ x0 Ay = 0 ∀ y ∈ Rr =⇒ x0 z = 0 ∀ z ∈ C(A) = S =⇒ x ∈ S ⊥ . ∴ N (A0 ) ⊆ S ⊥ and we have N (A0 ) = S ⊥ . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 7 / 22 From Theorem A.1, dim(N (A0 )) = n − rank(A0 ) = n − rank(A) = n − r. Therefore, dim(S) + dim(S ⊥ ) = dim(C(A)) + dim(N (A0 )) = r+n−r = n. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 8 / 22 Result A.4: Suppose a vector space S ⊆ Rn . Then any x ∈ Rn can be written as x = s + t, where s ∈ S, t ∈ S ⊥ . Furthermore, the decomposition is unique. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 9 / 22 Proof of Result A.4: Let a1 , . . . , ar be a basis for S. Let b1 , . . . , bk be a basis for S ⊥ . We know that r + k = n. We now show that a1 , . . . , ar , b1 , . . . , bk is a set of LI vectors and is thus a basis for Rn by V4. Suppose c1 a1 + · · · + cr ar + cr+1 b1 + · · · + cn bk = 0. Then c1 a1 + · · · + cr ar = −(cr+1 b1 + · · · + cn bk ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 10 / 22 Moreover, c1 a1 + · · · + cr ar ∈ S and − (cr+1 b1 + · · · + cn bk ) ∈ S ⊥ . Because these two vectors are equal each other, they are in S and S ⊥ and must be 0 ∵ S ∩ S ⊥ = {0}. By LI of a1 , . . . , ar , we have c1 = · · · = cr = 0. By LI of b1 , . . . , bk , we have cr+1 = · · · = cn = 0. Thus, a1 , . . . , ar , b1 , . . . , bk are LI and a basis for Rn . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 11 / 22 This implies that any x ∈ Rn may be written as x = d1 a1 + · · · + dr ar + dr+1 b1 + · · · + dn bk for some d1 , . . . , dn ∈ R. If we take s = d1 a1 + · · · + dr ar and t = dr+1 b1 + · · · + dn bk , we have s ∈ S, t ∈ S ⊥ and x = s + t. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 12 / 22 To prove the uniqueness, suppose x = s1 + t1 = s2 + t2 , where s1 , s2 ∈ S and t1 , t2 ∈ S ⊥ . Now show that s1 = s2 and t1 = t2 to complete the proof. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 13 / 22 s1 + t1 = s2 + t2 =⇒ s1 − s2 = t2 − t1 . Because S and S ⊥ are vectors spaces, s1 − s2 ∈ S, t2 − t1 ∈ S ⊥ . The equality of these vectors implies that s1 − s2 = t2 − t1 ∈ S ∩ S ⊥ =⇒ s1 − s2 = t2 − t1 = 0. Therefore, s1 = s2 , t2 = t1 . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 14 / 22 Suppose A = Suppose A = " # 1 0 " # 1 1 . Find and sketch C(A) and N (A0 ). . Find and sketch C(A) and N (A0 ). c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 15 / 22 If A = " # 1 0 , then C(A) = {x ∈ R2 : x2 = 0}. N (A0 ) = {x ∈ R2 : [1, 0]x = 0} = {x ∈ R2 : x1 = 0}. c Copyright 2012 Dan Nettleton (Iowa State University) X2 N(A') s+t t C(A) s X1 Statistics 611 16 / 22 If A = " # 1 1 , then C(A) = {x ∈ R2 : x1 = x2 }. N (A0 ) = {x ∈ R2 : [1, 1]x = 0} = {x ∈ R2 : x1 + x2 = 0}. c Copyright 2012 Dan Nettleton (Iowa State University) X2 C(A) s s+t X1 t N(A') Statistics 611 17 / 22 Result A.5: If A is an m × n matrix, then C(A)⊥ = N (A0 ). Proof of Result A.5: We essentially made this argument in our proof that dim(S) + dim(S ⊥ ) = n if S ∈ Rn is a vector space. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 18 / 22 Result A.6: Suppose S1 , S2 are vector spaces in Rn such that S1 ⊆ S2 . Then S2⊥ ⊆ S1⊥ . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 19 / 22 Proof of Result A.6: Suppose x ∈ S2⊥ . Then x0 y = 0 ∀ y ∈ S2 =⇒ x0 y = 0 ∀ y ∈ S1 =⇒ x ∈ S1⊥ . We have S2⊥ ⊆ S1⊥ . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 20 / 22 Suppose S is a vector space in Rn . Prove that S = (S ⊥ )⊥ . c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 21 / 22 Proof x ∈ S =⇒ x0 y = 0 ∀ y ∈ S ⊥ =⇒ x ∈ (S ⊥ )⊥ . ∴ S ⊆ (S ⊥ )⊥ . Now suppose x ∈ (S ⊥ )⊥ . Then x0 y = 0 ∀ y ∈ S ⊥ . By Result A.4, x = s + t for some s ∈ S and some t ∈ S ⊥ . By the previous two points, 0 = x0 t = (s + t)0 t = t0 t. ∴ t = 0 and x = s ∈ S so that (S ⊥ )⊥ ⊆ S. c Copyright 2012 Dan Nettleton (Iowa State University) Statistics 611 22 / 22