Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Generating a d-dimensional linear subspace efficiently Raphael Yuster SODA’10 Problem statement • Computing the rank of a matrix, and, more generally, generating a row-reduced equivalent matrix is a fundamental computational algebraic problem. • Standard approach: Gaussian elimination. This takes O(nω) < O(n2.38) time [H-B 74’, I-M-H 82’] . • No faster algorithm is known for computing the rank of general n n matrices. • If matrix is sparse (only m << n2 non-zero entries), can we compute a row reduced matrix faster? Not known (and probably impossible via G.E.) 2 Problem statement and main result • If we allow randomness, the rank can be computed faster. O(m n) Monte Carlo [Wiedemann 86’]. O(n3-1/(ω -1)) if m=O(n) Las Vegas [Eberly et al. 07’]. • Recall: If B is a row-reduced matrix of A then: B spans Rows(A) . B has rank(A) rows. Any d rows of B span a d-dim. subspace of Rows(A) . • Our main result: generating a d-dimensional subspace of Rows(A) requires (in some cases significantly) less time than computing a row-reduced matrix. 3 Main result • Let A be an n n over an arbitrary field, with m nonzero entries, and let d n: There is an algorithm that computes a d n matrix with full row rank which is a subspace of Rows(A). If rank(A) < d the algorithm generates a rank(A) n row-reduced matrix of A. Dense case • The running time is: Sparse case O(min{ n2-2/ω m1/ωdω-2+1/ω , n2dω-2 } ) . 4 Implications and comparison • Can be used to answer queries of the form “rank(A) > d ?” • A row-reduced matrix of A, and thus rank(A), computable in time which is a function of the rank (binary search). O(min{ n2-2/ω m1/ω rank(A)ω-2+1/ω , n2 rank(A)ω-2 } ) • If the matrix is dense, the runtime is O(n2dω-2) < O(n2d0.38). For d = o(n) this is faster than doing G.E. from scratch. • Even more efficiency if matrix is sparse: 5 Implications and comparison • If m = n2-ε then for d suff. small we get better than O(n2dω-2) and sometimes even better than O(n2) : d < (n2/m)0.528 < (n2/m) 1/(ω-1)2 d < (n2/m) sub-quadratic runtime. running time O(n2dω-2 (md/n2)1/ω). o(1) 6 An application • Given a graph G with n vertices what’s the cardinality M(G) of the maximum matching? • [Lovász 79’] proved that computing 2M(G) (randomly) reduces to computing the rank of an n n matrix over a small finite field. • Thus, his algorithm computes M(G) in O(nω) time. • Our algorithm computes M(G) in O(n2 M(G)ω-2) time, which is faster whenever M(G) = o(n) . 7 Proof of main result - outline Proof consists of two parts: 1. Reducing the computation of a d-dim. subspace of a sparse matrix to the computation of a d-dim. subspace of a dense, but smaller, rectangular matrix. 2. Computing a d-dim. subspace of a (possibly dense, possibly rectangular) matrix by repeatedly filtering out non-essential rows and replacing other rows with a smaller equivalent set of rows spanning the same subspace. 8 Proof of main result – part 1 We need a result of Ibarra, Moran and Hui 82’: Lemma 1: Let B be an l n matrix. Then a row-reduced matrix can be computed in O(nl ω-1) time. We prove: Sparse Lemma: Let A be an n n matrix with m nonzeroes. Let d n2/m. For any s with d s n there is a matrix L with at most dn/s rows so that: • If rank(A) < d then L is row-equivalent to A. • If rank(A) d then Rows(L) is a subspace of dimension at least d of Rows(A). • L is constructed in O(msω-1) time. 9 Proof of main result – part 1 Proof of Lemma 2: Partition the rows of A into t = n/s rectangular matrices, A1,…,At , each with s rows. Let ci denote the number of non-zero columns of Ai. t c m i 1 i Compact Ai by discarding zero columns (but remember their locations) thereby obtaining an s ci matrix Bi. Apply Lemma 1 to each Bi to obtain reduced equivalent matrices Ci with ci columns and rank(Bi)=rank(Ai) rows. 10 Proof of main result – part 1 The time required to construct Ci is O(cisω-1). So, to construct all C1,…,Ct takes O(msω-1). Let Di be the matrix obtained from Ci by taking the first d rows of Ci. Expand Di by reinserting the zero columns. If rank(Ai) < d then Di is row-equivalent to Ai Otherwise, Di spanning a d-dim. subspace of Rows(Ai). 11 Proof of main result – part 1 A1 A2 A3 B1 4 0 1 0 3 0 4 4 0 9 0 3 0 4 0 6 8 2 5 0 2 0 1 0 2 4 2 0 0 8 3 0 0 4 0 0 0 8 0 8 0 0 0 1 0 0 2 1 3 0 0 6 3 0 8 0 0 1 0 2 1 0 4 4 1 3 4 4 0 9 0 5 0 3 6 9 0 0 0 3/4 -11/4 -3 -1 2 0 0 7 9 4 0 1 0 7 4 1 3 4 4 0 9 3 4 6 8 2 5 2 1 2 4 2 C1 D1 4 0 1 0 3 0 4 4 0 0 0 3/4 0 -11/4 0 -3 -1 2 12 Proof of main result – part 1 Let L be the union of the Di. L has at most dt = dn/s rows. If rank(A) < d then L is equivalent to A, as each Di is equivalent to Ai. If rank(A) d then Rows(L) is a subspace of Rows(A) with dimension at least d. 13 Proof of main result – part 2 Dense Lemma: Let L be an l n matrix and let d l /2. We construct L* with at most l /2 rows so that: If rank(L) < d then L* is row-equivalent to L. If rank(L) d then Rows(L*) is a subspace of Rows(L) with dimension at least d. The running time is O(lnd ω-2). 14 Proof of main result – part 2 Proof of the main result – the algorithm: We are given an n n matrix A containing m nonzero entries, and a positive integer d. •If d < n2/m apply the sparse lemma with a value s that will be chosen later after optimization. The result is a matrix L0 with n column and l0 dn/s rows satisfying the conditions of the lemma. •If d n2/m we just set L0=A and l0= n. In any case now Rows(L0) is a subspace of Rows(A) with dimension at least d (or else L0 is equivalent to A). 15 Proof of main result – part 2 We check if d l0 /2. If so, we apply the dense lemma to L0 and obtain L1 with l1 l0 /2 rows. We repeatedly apply the dense lemma to Li as long as d li /2 and obtain the next matrix Li +1. We halt when d > li /2. After halting, we perform a final Gaussian elimination on Li and obtain a reduced matrix B, which is the output of our algorithm. 16 Proof of main result – part 2 Proof of the main result- correctness: At any stage of the algorithm, the rows of the current matrix Li form a subspace of Rows(A), and hence also at the end Rows(B) is a subspace of Rows(A). Thus, if B has (at least) d rows then the final output is (at least) a d-dimensional subspace of Rows(A). If Rank(A) < d then each of the Li, as well as the final B, is row-equivalent to A. 17 Proof of main result – part 2 Proof of the main result- running time: •Consider first the case when d n2/m. In this case l0= n and the application of the dense lemma on Li requires O(2-in2d ω-2) time. After O(log n) applications the final G.E. to obtain B takes O(nd ω-1) time by Lemma 1. The overall runtime is O(n2d ω-2) as required. •Consider next the case when d < n2/m. The application of the sparse lemma requires O(msω-1) time. 18 Proof of main result – part 2 As l0 dn/s, the dense lemma on Li requires: O(2-is-1n2d ω-1) time. After O(log n) applications the final G.E. to obtain B takes O(nd ω-1) time by Lemma 1. Thus, the overall running time is O(msω-1 + s-1n2d ω-1) Choosing optimal value for s which is s = (d ω-1n2/m)1/ω we obtain a runtime of O(n2-2/ω m1/ωdω-2+1/ω) as required. 19 Thanks 20