* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download The main theorem
Vector space wikipedia , lookup
Shapley–Folkman lemma wikipedia , lookup
System of linear equations wikipedia , lookup
Linear least squares (mathematics) wikipedia , lookup
Capelli's identity wikipedia , lookup
Rotation matrix wikipedia , lookup
Determinant wikipedia , lookup
Non-negative matrix factorization wikipedia , lookup
Gaussian elimination wikipedia , lookup
Matrix (mathematics) wikipedia , lookup
Principal component analysis wikipedia , lookup
Four-vector wikipedia , lookup
Matrix calculus wikipedia , lookup
Symmetric cone wikipedia , lookup
Jordan normal form wikipedia , lookup
Singular-value decomposition wikipedia , lookup
Eigenvalues and eigenvectors wikipedia , lookup
Matrix multiplication wikipedia , lookup
Perron–Frobenius theorem wikipedia , lookup
30 The key fact for dealing with two or more matrices in A is that A is commutative; that is, if M ∈ A and N ∈ A then MN = NM. This follows from Lemma 1.2. The following rather technical lemma is the building block for the main result. Lemma 2.4 Let U1 , . . . , Um be non-zero orthogonal subspaces of RΩ such that RΩ = U1 ⊕ · · · ⊕ Um . For 1 6 i 6 m, let Qi be the orthogonal projector onto Ui . Let M be a symmetric matrix which commutes with Qi for 1 6 i 6 m. Let the eigenspaces of M be V1 , . . . , Vr . Let the non-zero subspaces among the intersections Ui ∩V j be W1 , . . . , Wt . Then (i) RΩ = W1 ⊕ · · · ⊕Wt ; (ii) if k 6= l then Wk ⊥ Wl ; (iii) for 1 6 k 6 t, the space Wk is contained in an eigenspace of M; (iv) for 1 6 k 6 t, the orthogonal projector onto Wk is a polynomial in Q1 , . . . , Qm and M. Proof Part (iii) is immediate from the definition of the Wk . So is part (ii), because Ui1 ⊥ Ui2 if i1 6= i2 and V j1 ⊥ V j2 if j1 6= j2 so the two spaces Ui1 ∩V j1 and Ui2 ∩V j2 are either orthogonal or equal. For 1 6 j 6 r, let Pj be the projector onto V j . Since Pj is a polynomial in M, it commutes with Qi , so Ui is geometrically orthogonal to V j . By Lemma 2.2, the orthogonal projector onto Ui ∩V j is Qi Pj , which is a polynomial in Qi and M. This proves (iv). Finally, Qi Pj = O precisely when Ui ∩ Pj = {0}, so the sum of the orthogonal projectors onto the spaces Wk is the sum of the non-zero products Qi Pj , which is equal to the sum of all the products Qi Pj . But ! ! ∑ ∑ Qi P j = ∑ Qi i j i ∑ Pj = I 2 = I, j so Lemma 2.3 shows that RΩ = W1 ⊕ · · · ⊕Wt , proving (i). Corollary 2.5 If M and N are commuting symmetric matrices in RΩ×Ω then RΩ is the direct sum of the non-zero intersections of the eigenspaces of M and N. Moreover, these spaces are mutually orthogonal and their orthogonal projectors are polynomials in M and N. Proof Apply Lemma 2.4 to M and the eigenspaces of N. If W is contained in an eigenspace of a matrix M, I shall call it a sub-eigenspace of M. 31 Theorem 2.6 Let A be the Bose-Mesner algebra of an association scheme on Ω with s associate classes and adjacency matrices A0 , A1 , . . . , As . Then RΩ has s +1 mutually orthogonal subspaces W0 , W1 , . . . , Ws , called strata, with orthogonal projectors S0 , S1 , . . . , Ss such that (i) RΩ = W0 ⊕W1 ⊕ · · · ⊕Ws ; (ii) each of W0 , W1 , . . . , Ws is a sub-eigenspace of every matrix in A ; (iii) for i = 0, 1, . . . , s, the adjacency matrix Ai is a linear combination of S0 , S1 , . . . , Ss ; (iv) for e = 0, 1, . . . , s, the stratum projector Se is a linear combination of A0 , A1 , . . . , As . Proof The adjacency matrices A1 , . . . , As commute and are symmetric, so s − 1 applications of Lemma 2.4, starting with the eigenspaces of A1 , give spaces W0 , . . . , Wr as the non-zero intersections of the eigenspaces of A1 , . . . , As , where r is as yet unknown. These spaces We are mutually orthogonal and satisfy (i). Since A0 = I and every matrix in A is a linear combination of A0 , A1 , . . . , As , the spaces We clearly satisfy (ii). Each Se is a polynomial in A1 , . . . , As , hence in A , so (iv) is satisfied. Let C(i, e) be the eigenvalue of Ai on We . Then Equation (2.2) shows that r Ai = ∑ C(i, e)Se e=0 and so (iii) is satisfied. Finally, (iii) shows that S0 , . . . , Sr span A . Suppose that there are real scalars λ0 , . . . , λr such that ∑e λe Se = O. Then for f = 0, . . . , r we have O = (∑e λe Se ) S f = λ f S f so λ f = 0. Hence S0 , . . . , Sr are linearly independent, so they form a basis for A . Thus r + 1 = dim A = s + 1 and r = s. We now have two bases for A : the adjacency matrices and the stratum projectors. The former are useful for addition, because A j (α, β) = 0 if (α, β) ∈ Ci and i 6= j. The stratum projectors make multiplication easy, because Se Se = Se and Se S f = O if e 6= f . Before calculating any eigenvalues, we note that if Ais the adjacency matrix of any subset C of Ω × Ω then Aχα = ∑ χβ : (β, α) ∈ C . In particular, Ai χα = χCi (α) and Jχα = χΩ . Furthermore, if M is any matrix in RΩ×Ω then MχΩ = ∑ω∈Ω Mχω . If M has constant row-sum r then MχΩ = rχΩ . 32 Example 2.2 Consider the groups of size k and n 1 A1 (α, β) = 0 n 1 A2 (α, β) = 0 group divisible association scheme GD(b, k) with b if α and β are in the same group but α 6= β otherwise if α and β are in different groups otherwise. Consider the 1-dimensional space W0 spanned by χΩ . We have A0 χΩ = IΩ χΩ = χΩ A1 χΩ = a1 χΩ = (k − 1)χΩ A2 χΩ = a2 χΩ = (b − 1)kχΩ . Thus W0 is a sub-eigenspace of every adjacency matrix. This does not prove that W0 is a stratum, because there might be other vectors which are also eigenvectors of the all the adjacency matrices with the same eigenvalues as W0 . Let α and β be in the same group ∆. Then A1 χα = χ∆ − χα A2 χα = χΩ − χ∆ A1 χβ = χ∆ − χβ A2 χβ = χΩ − χ∆ so A1 (χα − χβ ) = −(χα − χβ ) and A2 (χα − χβ ) = 0. Let Wwithin be the b(k − 1)dimensional “within groups” subspace spanned by all vectors of the form χα − χβ with α and β in the same group. Then Wwithin is a sub-eigenspace of A0 , A1 and A2 and the eigenvalues for A1 and A2 are different from those for W0 . Since eigenspaces are mutually orthogonal, it is natural to look at the orthogonal complement of W0 +Wwithin . This is the (b−1)-dimensional “between groups” subspace Wbetween spanned by vectors of the form χ∆ − χΓ where ∆ and Γ are different groups. Now A1 χ∆ = A1 and ∑ χα = kχ∆ − ∑ χα = (k − 1)χ∆ α∈∆ A2 χ∆ = A2 α∈∆ ∑ χα = k(χΩ − χ∆) α∈∆ so A1 (χ∆ −χΓ ) = (k −1)(χ∆ −χΓ ) and A2 (χ∆ −χΓ ) = −k(χ∆ −χΓ ). Thus Wbetween is a sub-eigenspace with different eigenvalues from Wwithin . Therefore the strata are W0 , Wwithin and Wbetween . 33 Lemma 2.7 If P ∈ A and P is idempotent then P = ∑e∈F Se for some subset F of {0, . . . , s}. Proof Let P = ∑se=0 λe Se . Then P2 = ∑se=0 λ2e Se , which is equal to P if and only if λe ∈ {0, 1} for e = 0, . . . , s. For this reason the stratum projectors are sometimes called minimal idempotents or primitive idempotents. Lemma 2.8 The space W spanned by χΩ is always a stratum. Its projector is |Ω|−1 JΩ . Proof The orthogonal projector onto W is |Ω|−1 JΩ because JΩ χΩ = and ∑ JΩχω = |Ω| χΩ ω∈Ω JΩ (χα − χβ ) = 0. This is an idempotent contained in A , so it is equal to ∑e∈F Se , for some subset F of {0, . . . , s}, by Lemma 2.7. Then 1 = dimW = tr |Ω|−1 JΩ = ∑ tr Se = ∑ dimWe e∈F e∈F so we must have |F | = 1 and W is itself a stratum. Notation The 1-dimensional stratum spanned by χΩ is always called W0 . Although there are the same number of strata as associate classes, there is usually no natural bijection between them. When I want to emphasize this, I shall use a set K to index the associate classes and a set E to index the strata. However, there are some association schemes for which E and K are naturally the same but for which W0 does not correspond to A0 . So the reader should interpret these two subscripts ‘0’ as different sorts of zero. I shall always write de for dimWe .