Abstract
The Gaussian elimination with partial pivoting (GEPP) is a classical algorithm for solving systems of linear equations. Although in specific cases the loss of precision in GEPP due to roundoff errors can be very significant, empirical evidence strongly suggests that for a typical square coefficient matrix, GEPP is numerically stable. We obtain a (partial) theoretical justification of this phenomenon by showing that, given the random \(n\times n\) standard Gaussian coefficient matrix A, the growth factor of the Gaussian elimination with partial pivoting is at most polynomially large in n with probability close to one. This implies that with probability close to one the number of bits of precision sufficient to solve \(Ax = b\) to m bits of accuracy using GEPP is \(m+O(\log n)\), which improves an earlier estimate \(m+O(\log ^2 n)\) of Sankar, and which we conjecture to be optimal by the order of magnitude. We further provide tail estimates of the growth factor which can be used to support the empirical observation that GEPP is more stable than the Gaussian Elimination with no pivoting.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
The Gaussian Elimination is a classical algorithm for solving systems of linear equations [6, Chapter 3], [7, Chapter 9]. The simplest form of the algorithm—the Gaussian Elimination with no pivoting—solves a linear system (SLE) \(Ax=b\) with a square coefficient matrix A by performing the LU–factorization: A is represented as the product LU where L and U are lower and upper triangular matrix, respectively, and x is obtained by a combination of forward and back substitutions \(y:=L^{-1}b\), \(x:=U^{-1}y\). A possible algorithmic representation of this well known process is given below in Algorithm 1. The procedure produces a sequence of matrices \(A^{(0)}:=A,A^{(1)},\dots ,A^{(n-1)}=:U\), where for every \(k\le n-1\), the \(k\times k\) top left submatrix of \(A^{(k-1)}\) is upper triangular. The elimination process with no pivoting fails if at any step \(k=1,\dots ,n-1\), the k–th diagonal element of \(A^{(k-1)}\) is zero.
The computation of \(A^{(k)}\) from \(A^{(k-1)}\) can be represented in matrix form as
where
and where \(e_1,\dots ,e_n\) are standard unit basis vectors in \({\mathbb {R}}^n\). We also note that with this notation
The matrices \(\textrm{Id}_n-\tau ^{(k)}e_k^\top \), \(k=1,\dots ,n-1\), are called the Gauss transformations.
When considering an implementation using the floating point arithmetic, a well known issue of the Gaussian Elimination is its numerical instability. Recall that the condition number \(\kappa (A)\) of a square matrix A defined as the ratio of the largest and smallest singular value of A. Even for some well-conditioned matrices (i.e having a small condition number), solving SLE with help of the Gaussian Elimination with no pivoting results in large relative errors of the computed solution vectors [6, Section 3.3].
Several modifications of the elimination procedure are commonly used in matrix computations to address the instability issue [6, Chapter 3], [7, Chapter 9]. In particular, the Gaussian Elimination with Partial Pivoting (GEPP) looks for a representation \(PA=LU\) (called PLU–factorization), where, as before, L and U are lower and upper triangular matrices, while P is a specially constructed permutation matrix. The solution of a corresponding SLE can then be obtained by a combination of forward and back substitutions, and a permutation of vector’s components (see Algorithm 2; for better readability, we represent the formula for L in matrix rather than entry-wise form there). The GEPP succeeds in exact arithmetic whenever A is non-singular (although it may fail in floating point arithmetic).
A seminal result of Wilkinson [21] gives an upper bound on the backward error during the Gaussian Elimination when the floating point computations are performed. Define the unit roundoff
so that for every real number x, its floating point representation \(\textrm{fl}(x)\) satisfies \(|x-\textrm{fl}(x)|\le \textbf{u}|x|\) as long as no underflow or overflow exception arise [6, Section 2.7]. Let A be an invertible \(n\times n\) matrix, assume that GEPP in floating point arithmetic with no underflow and overflow exceptions is performed on the matrix \(\textrm{fl}(A)\), and assume that no error occurs during the computation (i.e no zero pivots are encountered). Let \({\hat{P}},{\hat{L}},{\hat{U}}\) be the computed matrices from the floating point GEPP of \(\textrm{fl}(A)\), with \({\hat{U}}_{n,n}\ne 0\), and let P, L, U be the PLU–factorization of A in exact arithmetic. Assume that \({\hat{P}}=P\). Further, let \({\hat{x}}\) denote the computed solution corresponding to the exact solution of the SLE \(Ax=b\). Then
where
and
(see, in particular, [6, Theorem 3.3.1 and Theorem 3.3.2]). Define the growth factor \(\mathbf{g_{\textrm{GEPP}}}\) as
where \({\hat{A}}^{(1)},\dots ,{\hat{A}}^{(n-1)}\) denote the computed (in the floating point arithmetic) matrices \(A^{(1)},\dots \), \(A^{(n-1)}\). Then, under the above assumptions, the backward error estimate can be written as
implying, under the additional assumption \(s_{\min }(A)\ge 2\Vert E\Vert \), the forward error bound for the computed solution
where \(\kappa (A)=\Vert A\Vert \,\Vert A^{-1}\Vert \) is the condition number of A. Similar error bounds are available for other versions of the Gaussian Elimination (with no pivoting, with complete or with rook pivoting). We refer, in particular, to Wilkinson’s paper [21] and to modern accounts of the backward error analysis of the different forms of the Gaussian Elimination in [6, Chapter 3], [7, Chapter 9], as well as [8].
It can be checked that \(\textbf{g}_{\textrm{GEPP}}(A)=O(2^n)\) for any \(n\times n\) invertible matrix A, and that this bound is attained. Thus, (2) provides a satisfactory worst-case estimate only under the assumption \(\textbf{u}\ll 2^{-n}\), i.e when the unit roundoff is exponentially small in the matrix dimension. At the same time, the accumulated empirical evidence suggests that for a “typical” coefficient matrix the loss of precision is much smaller than the worst-case prediction. Let us quote [6, p. 131]: “Although there is still more to understand about [the growth factor], the consensus is that serious element growth in Gaussian Elimination with Partial Pivoting is extremely rare. The method can be used with confidence.”
In [18] Trefethen and Schreiber carried out an empirical study of the Gaussian Elimination with Partial and with Complete Pivoting in the setting when the input coefficient matrix A is random, having i.i.d standard Gaussian entries. Their experiments showed that with high probability the growth factor in GEPP is only polynomially large in n. Further numerical studies by Edelman suggest that \(\textbf{g}_{\textrm{GEPP}}(A)\) of an \(n\times n\) standard Gaussian matrix A is of order \(O(n^{1/2+o(1)})\) with probability close to one (see a remark in [5, p. 182]).
An important step in improving theoretical understanding of numerical stability of the Gaussian Elimination was made by Yeung and Chan [22]. Their result implies (although that is not explicitly stated in the paper) that for the Gaussian Elimination with no pivoting applied to the standard \(n\times n\) Gaussian matrix, the relative error of the solution vector can be bounded above by \(\textbf{u}\,n^{O(1)}\) with probability close to one. A vast generalization of their estimate was obtained by Sankar ET AL. [15] in the context of the smoothed analysis of algorithms. Let M be any non-random \(n\times n\) matrix, and let G be an \(n\times n\) matrix with i.i.d \(N(0,\sigma ^2)\) Gaussian entries. The main result of [15] asserts that the expected number of bits of precision sufficient to solve \((M+G)x = b\) to m bits of accuracy using Gaussian elimination without pivoting is at most \(m+O\big (\log \big (n+\frac{\Vert M\Vert }{\sigma }\big )\big )\). This provides a theoretical justification for the observed performance of the GE with no pivoting for structured dense coefficient matrices.
The no-pivoting strategy is crucial for the proofs in [22] or [15]. With partial pivoting, the permutations of the rows after each elimination step introduce complex dependencies to the model which require other arguments to handle. In the PhD thesis [14], Sankar carried out smoothed analysis of GEPP based on certain recursive matrix formula (to be discussed in some detail in the next section). Let \(A=M+G\), where G is the Gaussian random matrix with i.i.d \(N(0,\sigma ^2)\) entries, and M is a deterministic matrix of spectral norm at most one. One of main results of [14] states that, with the above notation,
so that in the mean zero setting \(M=0\), with high probability \(\frac{\max _{k,i,j \in [n]} |A^{(k-1)}_{i,j}|}{\max _{i,j\in [n]} |A_{i,j}|}= n^{O(\log n)}\). Note that the quantity considered in [14] is not a growth factor as was defined above but its “exact arithmetic” counterpart. The relation between matrices \(A^{(k-1)}\) and the corresponding computed matrices \({\hat{A}}^{(k-1)}\) is not trivial and will be discussed later; at this point we note that assuming that the magnitudes of the ratio \(\frac{\max _{k,i,j \in [n]} |A^{(k-1)}_{i,j}|}{\max _{i,j\in [n]} |A_{i,j}|}\) and the growth factor \(\textbf{g}_{\textrm{GEPP}}(A)\) match and in view of (2), the result of Sankar implies that with high probability GEPP results in at most \(O(\log ^2 n)\) lost bits of precision in the obtained solution vector. This bound is worse than the \(O(\log n)\) estimate for GE with no pivoting implied by [22].
To summarize, whereas strong results on average-case stability of GE with no pivoting has been obtained in the literature, the Gaussian Elimination with Partial Pivoting lacked matching theoretical guarantees, let alone justifying the common belief that GEPP tends to be more stable than GE with no pivoting. In this work, we make progress on this problem. To avoid any ambiguity, we recall all the imposed assumptions and notation:
Theorem A
There are universal constants \(C,{\tilde{C}}>1\) and a function \(\tilde{n}:[1,\infty )\rightarrow {\mathbb {N}}\) with the following property. Let \(p\ge 1\), and let \(n\ge {\tilde{n}}(p)\).
-
Assume that the floating point computations with no underflow and overflow exceptions and with a unit roundoff \(\textbf{u}\) are being performed.
-
Let A be the random \(n\times n\) matrix with i.i.d standard Gaussian entries, (the real Ginibre Ensemble). Assume that the floating point GEPP is performed on the matrix \(\textrm{fl}(A)\).
Then with probability at least \(1-\textbf{u}^{1/8}\,n^{{\tilde{C}}}\), the GEPP for \(\textrm{fl}(A)\) succeeds in floating point arithmetic and the computed permutation matrix \({\hat{P}}\) agrees with the matrix P from the PLU–factorization of A in exact arithmetic. Furthermore, assuming \(\textbf{u}^{1/8}\,n^{{\tilde{C}}}\le 1/2\),
We do not attempt to compute the constant C in the above theorem explicitly and leave the problem of finding an optimal (up to \(n^{o_n(1)}\) multiple) estimate of the growth factor \(\textbf{g}_\textrm{GEPP}(A)\) for future research (see Sect. 8). Further, we expect a much stronger bound on the probability that GEPP succeeds in the floating point arithmetic and that \({\hat{P}}=P\).
In view of the aforementioned work of Wilkinson and well known estimates for the condition number of the Gaussian matrix [4, 17], the theorem implies that with probability close to one the number of bits of precision sufficient to solve \(Ax = b\) to m bits of accuracy using GEPP is \(m+O(\log n)\). We conjecture that this bound is optimal in the sense that in the same setting \(m+\Omega (\log n)\) bits of precision are necessary with probability close to one.
Let us further apply Theorem A to compare numerical stability of GEPP with that of GE with no pivoting. As we mentioned at the beginning of the introduction, the Gaussian Elimination with no pivoting can produce arbitrarily large relative error in the floating point arithmetic even for well-conditioned coefficient matrices. As an illustration, consider a \(2\times 2\) standard Gaussian matrix in floating point arithmetic,
The Gaussian Elimination with no pivoting yields the computed LU–factorization of M,
It can be checked that for every \(\varepsilon \in (\textbf{u},1)\), with probability \(\Theta (\varepsilon )\) all of the following holds:
-
The matrix M is well-conditioned, say, \(\kappa (M)\le 100\);
-
\(|\textrm{fl}(g_{11})|\le \varepsilon \), \(|\textrm{fl}(g_{12})|,|\textrm{fl}(g_{21})|,|\textrm{fl}(g_{22})|\in [1/2,2]\);
-
\(\big |\textrm{fl}\big (\textrm{fl}(g_{22})-\textrm{fl}(g_{12})\cdot \textrm{fl}(g_{21})/\textrm{fl}(g_{11})\big ) -\big (\textrm{fl}(g_{22})-\textrm{fl}(g_{12})\cdot \textrm{fl}(g_{21})/\textrm{fl}(g_{11})\big ) \big |\ge \Omega (\textbf{u})\,\big (\textrm{fl}(g_{22})-\textrm{fl}(g_{12})\cdot \textrm{fl}(g_{21})/\textrm{fl}(g_{11})\big )\).
With the above conditions, the bottom right element of the product \({\hat{L}}{\hat{U}}\) differs from \(\textrm{fl}(g_{22})\) by a quantity of order \(\Omega (\textbf{u})\,\big (\textrm{fl}(g_{22})-\textrm{fl}(g_{12})\cdot \textrm{fl}(g_{21})/\textrm{fl}(g_{11})\big )\), that is, the normwise backward error satisfies
for some universal constant \(c>0\) (one may safely take \(c=1/100\), say).
In sharp contrast with the above observation, in the case of GEPP the probability of large deviations for the backward error is much smaller as Theorem A shows. Indeed, with the notation from the theorem and in view of Wilkinson’s bound, arbitrary \(p\ge 1\) and assuming n is sufficiently large, we have
for a universal constant \(C'>0\). Thus, the tail of the distribution of the backward error of GEPP decays superpolynomially. Informally, the “proportion” of well-conditioned coefficient matrices yielding large backward errors is much smaller for GEPP than for the Gaussian Elimination with no pivoting.
We provide a detailed outline of the argument, as well as a comparison of our techniques with the earlier approach of Sankar, in the next section.
The following notation will be used throughout the paper:
For positive integers \(m\le n\),
- [n]:
-
is the set \(\{1,2,3,\dots , n\}\)
- [m, n]:
-
is the set \(\{m,m+1,\dots , n\}\)
For a \(n\times m\) matrix M, indices \(i\in [n]\), \(j\in [m]\), and non-empty subsets \(I \subset [n]\) and \(J \subset [m]\),
- \(M_{I,J}\):
-
is the submatrix of M formed by taking rows indexed over I and columns indexed over J. When \(I =\{i\}\) or \(J=\{j\}\), we will use lighter notations \(M_{i,J}\) and \(M_{I,j}\) in place of \(M_{\{i\},J}\) and \(M_{I,\{j\}}\)
- \(M_{i,j}\):
-
is the (i, j)th entry of M
- \(s_j(M)\):
-
is the jth largest singular value of M
- \({\mathbb {R}}^I\):
-
The |I|-dimensional Euclidean space with components indexed over I
- \(\textrm{dist}(\cdot ,\cdot )\):
-
The Euclidean distance
2 Outline of the proof
Let A be an \(n\times n\) standard Gaussian matrix, let \(A^{(0)}:=A,A^{(1)},\dots ,A^{(n-1)}\) be the sequence of matrices generated by GEPP process, and let \(P^{(1)},\dots ,P^{(n-1)}\) be the corresponding permutation matrices (see Algorithm 2). It turns out that in our probabilistic model, estimating the growth factor \(\textbf{g}_{\textrm{GEPP}}(A)\) can be reduced to bounding the exact arithmetic counterpart of the quantity,
Our main focus is to derive Proposition 6.13, which is the exact arithmetic counterpart of the main theorem, and then reduce the setting of floating-point arithmetic to exact arithmetic. We provide a rigorous account of the reduction procedure in Sect. 7, and prefer to avoid discussing this technical matter here. We only note that comparison of the matrices \(A^{(k-1)}\) and \({\hat{A}}^{(k-1)}\), \(1\le k\le n-1\), is based on a well established inductive argument somewhat similar to the one used to prove Wilkinson’s backward error bound. From now on and till Sect. 7 we work in exact arithmetic unless explicitly stated otherwise.
Define “unpermuted” matrices \({\mathcal {M}}^{(k)}\) obtained at the kth elimination step, i.e \({\mathcal {M}}^{(0)}:=A\) and
Let \(I_0:=\emptyset \), and for each \(1\le k\le n-1\) let \(I_k=I_k(A)\) be the (random) subset of [n] of row indices of A corresponding to the pivot elements used in the first k steps of the “permutation-free” elimination process. Notice that within the kth column of \({{\mathcal {M}}}^{(k)}\), the components except those in \(I_{k-1}\) and the kth pivot element, are all zeros. Therefore, the set \(I_k\) can be defined as
where \(\big \{i\in [n]{\setminus } I_{k-1}:\;{\mathcal {M}}^{(k)}_{ik}\ne 0\big \}\) is a singleton. We will further denote by \(i_k=i_k(A)\), \(1\le k\le n-1\), the elements in the singletons \(I_k{\setminus } I_{k-1}\), so that \(I_k = \{i_1,i_2,\dots , i_k \}\), \(1\le k\le n-1\).
For \(1\le k\le n-1\) and \(t \in [k]\), the first \(t-1\) components of \(\textrm{row}_{i_t}({\mathcal {M}}^{(k)})\) are zeros, and \({\mathcal {M}}^{(k)}_{ [n]\backslash I_k, [k] }\) is the zero matrix; more specifically, for each \(1 \le k \le n-1\), and \(j \in [n]\backslash I_k\),
(see, in particular, [15, Formula 4.1] for GE with no pivoting, which can be adapted to our setting). Thus, for \(0\le k<n-1\), the index \(i_{k+1}\) is defined as the one corresponding to the largest number among
Due to strong concentration of Gaussian variables, the operator norms of matrices \(A_{I,J}\), \(I,J\subset [n]\), can be uniformly bounded from above by a polynomial in n. Thus, the principal difficulty in obtaining satisfactory upper bounds on the growth factor \(\textbf{g}_{\textrm{GEPP}}(A)\) is in estimating the norm of vectors \(\textrm{row}_j(A_{[n],[k]}) \big (A_{I_k,[k]} \big )^{-1}\), \(j \in [n]\backslash I_k\). The sets \(I_k\) are random and depend on A in a rather complicated way. At the same time, the trivial upper bound
which completely eliminates the randomness of \(I_k\) from consideration, is vastly suboptimal.Footnote 1
The first part of this section is devoted to the argument of Sankar from [14] which yields a bound \(\textbf{g}_\textrm{GEPP}(A)=O(n^{C\log n})\) with high probability using certain recursive matrix formula. In the second part, we discuss our approach.
2.1 Sankar’s argument
Consider a block matrix
where \(B_{\mathrm{u\ell }}\) and \(B_{\mathrm{\ell r}}\) are square non-singular matrices and \(X\) is a row vector. Then, denoting \(B':= B_{\mathrm{\ell r}} - B_{\mathrm{\ell \ell }}B_{\mathrm{u\ell }}^{-1} B_{\textrm{ur}}\) and \(X':= X_{\textrm{r}} - X_{\mathrm{\ell }} B_{\mathrm{u\ell }}^{-1} B_\textrm{ur}\),
where \(B_{\textrm{u}}^\dagger \) is the right pseudoinverse of \(B_{\textrm{u}}\) (see [14, Chapter 3]).
The above formula is applied in [14] in a recursive manner. Assume for simplicity of exposition that we are interested in bounding the Euclidean norm of the vector \(A _{j,[n/2]} (A_{I_{n/2},[n/2]})^{-1}\) for some \(j\in [n]{\setminus } I_{n/2}\) (recall that, in view of (4) and standard concentration estimates for the spectral norm of Gaussian matrices, this would immediately imply an estimate on the components of \({\mathcal {M}}^{(n/2)}_{j,[n/2+1,n]}\)). Fix for a moment any \(0\le v< m < n/2\), and let \(B:= {\mathcal {M}}^{(v)}_{ I_{n/2}{\setminus } I_v, [v+1,n/2]}\) and \(X:= {\mathcal {M}}^{(v)}_{j, [v+1,n/2]}\). We write
It can be checked that with the above notation, \(B'= {\mathcal {M}}^{(m)}_{I_{n/2}{\setminus } I_m,[m+1,n/2]}\) and \(X' = {\mathcal {M}}^{(m)}_{j,[m+1,n/2]}\) [14]. Relation (5) then implies
Now, assume that we have constructed a sequence of indices \(0=k_0<k_1<k_2<\cdots k_s <n/2\), with \(k_s=n/2-1\) and \(k_1\ge n/4\). Applying the last relation recursively s times, we obtain
where by the definition of the partial pivoting, \(\big |{\mathcal {M}}^{(n/2-1)}_{j,n/2} ({\mathcal {M}}^{(n/2-1)}_{i_{n/2},n/2})^{-1}\big |\le 1\). Therefore, the problem reduces to estimating the spectral norms of matrices
Sankar shows that as long as \(n/2-k_\ell \) (\(\ell =s,s-1,\dots ,0\)) grow as a geometric sequence (in which case s should be of order logarithmic in n), the norm of each matrix can be bounded by a constant power of n with a large probability. We only sketch this part of the argument. Fix any \(0\le \ell <s\), and define \(Z:= (A_{I_{k_\ell },[k_\ell +1,n/2]})^\dagger A_{I_{k_\ell },[k_\ell ]}\), so that \(ZZ^\dagger = \textrm{Id}\), and
where, in view of (4),
Since \(Z^{\dagger }Z\) is a projection and has unit norm, an upper bound on \(\big \Vert (A_{I_{k_\ell },[k_\ell +1,n/2]})^\dagger \big \Vert \) would provide a bound on \(\big \Vert {\mathcal {M}}^{(k_\ell )}_{ (I_{n/2}{\setminus } I_{k_{\ell +1}})\cup \{j\}, [k_\ell +1,n/2]} Z \big \Vert \). The key observation here is that \(A_{I_{k_\ell },[k_\ell +1,n/2]}\) is equidistributed with the standard \(k_\ell \times (n/2-k_\ell )\) Gaussian matrix, so that a satisfactory estimate on the norm of the pseudoinverse follows.
Bounding the operator norm of \(\big ( {\mathcal {M}}^{(k_\ell )}_{I_{k_{\ell +1}}{\setminus } I_{k_\ell }, [k_\ell +1,n/2]} Z\big )^\dagger \) is more involved. Note that, equivalently, it is sufficient to provide a good lower bound on the smallest singular value of the matrix
We have
where, again in view of (4), for each admissible \(J\),
and where \(Z^{\dagger }Z\) is a \(k_\ell \times k_\ell \) orthogonal projection matrix of rank \(n/2-k_\ell \). Although \((A_{J,[k_\ell ]})^\top \) is dependent on Z, it can be shown that \(Z^{\dagger }Z (A_{J,[k_\ell ]})^\top \) behaves “almost” like \(Z^{\dagger }Z\) applied to an independent tall rectangular \(k_\ell \times (k_{\ell +1}-k_\ell )\) Gaussian matrix (see [14, Section 3.7]). This allows to obtain probabilistic estimates on the smallest singular value of the matrix in (9) which, under the assumption that the sequence \(n/2-k_\ell \) (\(\ell =s,s-1,\dots ,0\)) does not grow too fast, turn out to be strong enough to survive the union bound in (8).
To summarize, the above argument gives a polynomial in n estimate for matrices in (7), where s is logarithmic in n. Thus, (6) implies a bound \(\Vert {\mathcal {M}}^{(n/2)}_{j,[n/2+1,n]}\Vert _2=n^{O(\log n)}\), \(j\in [n]{\setminus } I_{n/2}\), with high probability. An extension of this argument to all \({\mathcal {M}}^{(k)}\), \(1\le k\le n-1\), yields \(\textrm{g}_{\textrm{GEPP}}(A)=n^{O(\log n)}\). As Sankar notes in [14], a different choice of s and of the sequence \(k_0,k_1,\dots ,k_s\), and a refined analysis for the operator norms of matrices (7) may improve the upper estimate on the growth factor, but cannot achieve a polynomial bound.
2.2 High-level structure of the proof of the main theorem
Returning to relation (4), a polynomial bound on the growth factor \(\textrm{g}_{\textrm{GEPP}}(A)\) will follow as long as the norm \((A_{I_r,[r]})^{-1}\) is bounded by \(n^{O(1)}\) for every \(1\le r\le n-1\) with high probability. We obtain this estimate via analysis of the entire singular spectrum of \(A_{I_r,[r]}\) rather than attempting to directly bound the smallest singular value of the matrix.
The strategy of the proof can be itemized as follows:
-
Obtaining estimates on the singular values of partially random block matrices. More specifically, we consider matrices of the form
$$\begin{aligned} B= \begin{bmatrix} F &{}\quad M \\ W &{}\quad Q \end{bmatrix}, \end{aligned}$$(10)where F is a fixed square matrix with prescribed singular spectrum, and M, W, Q are independent Gaussian random matrices of compatible dimensions. Our goal here is to derive lower bounds on the intermediate singular values of B in terms of singular values of F.
-
Applying the estimates on the intermediate singular values of partially random block matrices in a recursive manner together with a union bound argument, derive lower bounds on the “smallish” singular values of matrices \(A_{I_r,[r]}\). Our argument at this step only allows to bound first to \((r-C)\)th singular value of the matrix for some large constant C.
-
Use the bound on \(s_{r-C}(A_{I_r,[r]})\) together with the information on the Euclidean distances from \(\textrm{row}_{i_\ell }(A_{I_r,[r]})\) to \(\mathrm{span\,}\{\textrm{row}_{i_j}(A_{I_r,[r]}),\;1\le j<\ell \}\), \(\ell =1,\dots ,r\) that can be extracted from the partial pivoting rule, to obtain polynomial in n lower bounds on \(s_{\min }(A_{I_r,[r]})\).
Below, we discuss each component in more detail.
Singular spectrum of partially random block matrices The partially random block matrices are treated in Sect. 3 of the paper. Consider a block matrix B of type (10), where F is a fixed \(r\times r\) matrix, M is \(r\times x\), W is \(x\times r\), Q is \(x\times x\) (with \(x\le r\)), and the entries of M, W, Q are mutually independent standard Gaussian variables. In view of rotational invariance of the Gaussian distribution, we can “replace” F with a diagonal matrix D with the same singular spectrum, and with its singular values on the main diagonal arranged in a non-decreasing order. We fix a small positive parameter \(\tilde{\varepsilon }>0\) and an integer \(i\ge 1\) such that \(\tilde{\varepsilon }(1+\tilde{\varepsilon })^{-i}r\ge 2\). Our goal at this point is to estimate from below the singular value
Having chosen a certain small threshold \(\tau >0\) (which is defined as a function of \(i,r,\tilde{\varepsilon }\), the singular spectrum of D, and some other parameters which we are not discussing here), our estimation strategy splits into two cases depending on whether the number \(\ell _{i+1}'\) of the singular values of D less than \(\tau \) is “small” or “large”. In the former case, the matrix D has a well controlled singular spectrum, and our goal is to show that attaching to it x rows and columns of standard Gaussians cannot deteriorate the singular values estimates. In the latter case, we show that by adding the Gaussian rows and columns we actually improve the control of the singular values, using that the top left \(\ell _{i+1}'\times \ell _{i+1}'\) corner of B is essentially a zero matrix. The main result of Sect. 3—Proposition 3.3—provides a probability estimate on the event that the ratio
is small assuming certain additional relations between the parameters \(x,r,\tilde{\varepsilon }\).
A recursive argument to bound \(s_{r-C}(A_{I_r,[r]})\) The treatment of the partially non-random block matrices allows us to solve the principal problem with estimating the singular spectrum of \(A_{I_r,[r]}\), namely, the complicated dependencies between A and the index set \(I_r\). As we mentioned before, simply bounding the kth smallest singular value of \(A_{I_r,[r]}\) by \(\min \limits _{I\subset [n],\,|I|=r}s_{r-k}(A_{I_r,[r]})\) produces an unsatisfactory estimate for small k. On the other hand, in view of strong concentration of intermediate singular values, already for \(k\gg \sqrt{n}\,\textrm{polylog}(n)\) (see Proposition 3.2) this straightforward union bound argument does work. In order to boost the union bound argument to smaller k, we avoid taking the union bound over all \(I\subset [n],\,|I|=r\) and instead condition on a realization of \(I_{r'}\) for certain \(r'<r\), so that the union bound over all \(I_{r'}\subset I\subset [n]\) of cardinality r runs over only \({n-r'\atopwithdelims ()r-r'}\) admissible subsets rather than \({n\atopwithdelims ()r}\) subsets. The two main issues with this approach are
-
first, we must have estimates for the singular spectrum of \(A_{I_{r'},[r']}\) in order to apply the results of Sect. 3 to obtain bounds for the singular values of \(A_{I_r,[r]}\), and,
-
second, conditioning on a realization of \(I_{r'}\) inevitably destroys Gaussianity and mutual independence of the entries of \(A_{[n]{\setminus } I_{r'},[r']}\).
The first issue is resolved through the inductive argument, when estimates on the spectrum of \(A_{I_{r'},[r']}\) obtained at the last induction step are used to control the singular spectrum of \(A_{I_r,[r]}\) at the next step. Of course, in this argument we must make sure that the total error accumulated throughout the induction process stays bounded by a constant power of n.
The second issue with probabilistic dependencies is resolved by observing that the partial pivoting “cuts” a not too large set of admissible values for the elements in \(A_{[n]{\setminus } I_{r'},[r']}\) i.e we can continue treating them as independent Gaussians up to a manageable loss in the resulting probability estimate after conditioning on a certain event of not-too-small probability. This problem is formally treated by studying the random polytopes \(K_{r'}(A)\subset {\mathbb {R}}^n\) defined in Sect. 4 as
where
By the nature of the partial pivoting process, any row of the submatrix \(A_{[n]{\setminus } I_r',[n]}\) necessarily lies within the polytope \(K_{r'}(A)\), and its distribution is a restriction of the standard Gaussian measure in \({\mathbb {R}}^n\) to \(K_{r'}(A)\) (see Sect. 4 for a rigorous description). After showing that the Gaussian measure of \(K_{r'}(A)\) is typically “not very small”, we can work with the rows of \(A_{[n]{\setminus } I_r',[n]}\) as if they were standard Gaussian vectors, up to conditioning on an event of a not very small probability. We remark here that Sankar’s work [14] uses random polytopes related to our construction.
Estimating the smallest singular value of \(A_{I_r,[r]}\) To simplify the discussion, we will only describe the idea of showing that with a “sufficiently high” probability, \(\big (s_{\min }(A_{I_r,[r]})\big )^{-1}=n^{O(1)}\), without considering computation of the moments of \(\big (s_{\min }(A_{I_r,[r]})\big )^{-1}\). As a corollary of the lower bound on \(s_{r'-C}(A_{I_{r'},[r']})\) obtained via the recursive argument, we get that with high probability, the inverse of the smallest singular value of the rectangular matrix \(A_{I_{r'},[r'+2{\tilde{C}}]}\) satisfies \(\big (s_{\min }(A_{I_{r'},[r'+2{\tilde{C}}]})\big )^{-1}=n^{O(1)}\), \(r'\in [{\tilde{C}}+1, n-2{\tilde{C}}]\), for some integer constant \(\tilde{C}>0\) (see Corollary 5.3). This corollary is a quantitative version of a rather general observation that, by adding at least \(\ell +1\) independent Gaussian rows to a fixed square matrix with at most \(\ell \) zero singular values, we get a rectangular matrix with a strictly positive \(s_{\min }\) almost surely.
Once a satisfactory bound on \(s_{\min }(A_{I_{r'},[r'+2{\tilde{C}}]})\), \(r'\in [{\tilde{C}}+1, n-2{\tilde{C}}]\), is obtained, we rely on the simple deterministic relation between the smallest singular value and distances to rowspaces: for every \(m\times k\) matrix Q,
where \(H_i(Q)\) denotes the subspace spanned by row vectors \(Q_{j,[m]}\) for \(j\ne i\). In our context, a strong probabilistic lower bound on \(\textrm{dist}( \mathrm{span\,}\{A_{i_t,[r]},\,1\le t<s\}, A_{i_s,[r]})\) (\(s\le r\)) guaranteed by the partial pivoting strategy, implies that with high probability for every \(t\in [r-2{\tilde{C}}]\),
(see proof of Proposition 6.9), and via the above deterministic relation to the singular values,
This, combined with some auxiliary arguments, implies the lower bound on \(s_{\min }(A_{I_r,[r]})\).
3 Intermediate singular values of partially random block matrices
We start with a preparatory material to deal with norms and intermediate singular values of random matrices. We first consider a standard deviation estimates for the Hilbert–Schmidt norm of a Gaussian random matrix; see, for example, [1]:
Theorem 3.1
Let G be an \(u\times t\) random matrix with i.i.d standard Gaussian entries. Then
where \(c>0\) is a universal constant.
The next proposition was proved in the special case of square random Gaussian matrices by Szarek [17]. In a much more general setting, similar results were obtained earlier by Nguyen [12]; his argument was later reused in [9] to get sharp small ball probability estimates for the condition number of a random square matrix.
Proposition 3.2
(Singular values of random matrices with continuous distributions) Let M be an \(u\times t\) (\(t\ge u\)) random matrix with i.i.d. standard Gaussian entries. Then
where \(c'\in (0,1]\) is a universal constant.
We provide a proof of the above proposition in the “Appendix”.
This section deals with a large number of parameters satisfying multiple constraints; we group those constraints into blocks for better readability. We have four “section-wide” scalar parameters:
The objective of the section is to study singular values of a block matrix of the form
where F is a fixed \(r\times r\) matrix with prescribed singular values, M is \(r\times x\), W is \(x\times r\), Q is \(x\times x\), and the entries of M, W, Q are mutually independent standard Gaussians. Let
Observe that the relation \(\tilde{\varepsilon }(1+\tilde{\varepsilon })^{-i}r\ge 2\) from (11) yields, for \(j \in [0,i-1]\),
which in turn implies that the sequence \((\varrho _j)_{j=0}^i\) is strictly increasing.
Next, let \(g(\cdot ):(r_j)_{j \in [i]}\rightarrow (0,\infty )\) be a strictly positive growth function satisfying
Now, we assume the matrix F satisfies
In this section, we deal with an arbitrary growth function satisfying the conditions; a specific choice of \(g(\cdot )\) will be made later in Sect. 5.
Our objective in this section is to derive the following proposition.
Proposition 3.3
There are universal constants \(c\in (0,1]\), \({\tilde{C}}\ge 1\) with the following property. Let
where F is a fixed \(r\times r\) matrix, M is \(r\times x\), W is \(x\times r\), Q is \(x\times x\), and the entries of M, W, Q are mutually independent standard Gaussians. Assume that parameters \(\tilde{\varepsilon }\in (0,1]\), \(h\in (0,1]\), r, x, and \(i\in {\mathbb {N}}\) satisfy
where \(c' \in (0,1]\) is the constant from Proposition 3.2. Further, assume (14) for the singular values of F, for a positive function \(g(\cdot )\) satisfying (13). Then with probability at least
we have
Note that if \(F=UDV\) is a singular values decomposition of F then, in view of rotational invariance of the Gaussian distribution,
where \(U^{-1}M\), \(WV^{-1}\), and Q have mutually independent standard Gaussian entries, and where the singular spectrum of B coincides with that of
We can assume without loss of generality that the diagonal elements (the singular values) of D are arranged in the non-decreasing order when moving from top left to bottom right corner. We will work with the singular spectrum of \(B'\) as it will allow to somewhat simplify the computations.
The specific goal is to estimate from below the singular value
in terms of \(g(\varrho _i )\). To have a better control on probability estimates, we introduce one more scalar parameter \(h\in (0,1]\) which will allow us to balance the precision of the estimate and the probability with which the estimate holds (the smaller h is, the less precise the estimate is and the stronger are probability bounds). We set
and let
Let us remark that \(\ell _{i+1}' \le r - \varrho _i\) since, by the intermediate singular values assumption (14) on F, we have \(s_{\varrho _i}(F)\ge \tau \).
Set
Our argument to control \(s_{\lfloor (1-(1+\tilde{\varepsilon })^{-i-1})(r+x)\rfloor }(B)\) splits into two parts depending on whether \(\ell _{i+1}'\) is “small” or “large”. In the former case (see Lemma 3.4), the matrix F (or D) has a well controlled singular spectrum, and our goal is to show that attaching to it x rows and columns of standard Gaussians cannot deteriorate the singular values estimates. In this setting, we completely ignore the first \(\ell _{i+1}'\) rows of \(B'\), and work with the matrix \(B'_{I\times [r+x]}\). In the latter case (see Lemma 3.5), we show that by adding the Gaussian rows and columns we actually improve the control of the singular values. The fact that the top right \(\ell _{i+1}'\times x\) corner of \(B'\) is a standard Gaussian matrix, plays a crucial role in this setting. The proof of Proposition 3.3 follows from Lemmas 3.4 and 3.5.
The high-level proof strategy for both Lemmas 3.4 and 3.5 is similar. We construct a (random) subspace H of \({\mathbb {R}}^{r+x}\) of dimension at least \((1-(1+\tilde{\varepsilon })^{-i-1})(r+x)\), designed in such a way that, under appropriate assumptions on the singular spectra of certain submatrices of \(U^{-1}M\), \(WV^{-1}\), and Q, \(\Vert B'v\Vert _2\) is large for every unit vector \(v\in H\). By the minimax formula for singular values,
The “appropriate assumptions” on the singular spectra are encapsulated in a good event \({\mathcal {E}}_{good}\) which, as we show, has a very large probability. In what follows, it will be convenient to use notation
We remark that for every \(j=1,2,\dots ,i\), by the same derivation as shown in (12),
Lemma 3.4
There exist universal constants \(c\in (0,1]\), \({\tilde{C}}\ge 1\) with the following property. Assume that \(i \in {\mathbb {N}}\), \(\tilde{\varepsilon }\in (0,1]\), \(h\in (0,1]\), r, and x satisfy the assumptions of Proposition 3.3, and assume additionally that
where \(\ell _{i+1}'\) is defined in (16). Denote
where \(c' \in (0,1]\) is the constant from Proposition 3.2 and where \(\tau \) is defined by (15). Then with probability at least
we have
Proof
Construction of subspace H Denote by \(X_1,X_2,\dots ,X_{r+x}\in {\mathbb {R}}^{r+x}\) an orthonormal basis of the right singular vectors of the matrix \(B_{I\times [r+x]}'\), measurable w.r.t the \(\sigma \)–field \(\sigma (B_{I\times [r+x]}')\), where \(X_j\) corresponds to \(s_j(B_{I\times [r+x]}')\), \(1\le j\le r+x\), and where I is defined by (17). Note that by interlacing properties of the singular values (see, for example [3]), we have
in particular, \(s_{r-\ell _{i+1}'}(B_{I\times [r+x]}')\ge \tau \) everywhere on the probability space.
Observe that, conditioned on \(\sigma (B_{[r]\times [r+x]}')\), the \(x\times \ell _{i+1}''\) matrix
has mutually independent standard Gaussian entries. Denote by \(e^{(i+1)}_q\), \(1\le q\le \min (\lfloor \tilde{\varepsilon }x\rfloor ,\ell _{i+1}'')\), a random orthonormal system of right singular vectors of \(Y^{(i+1)}\) corresponding to \(\min (\lfloor \tilde{\varepsilon }x\rfloor ,\ell _{i+1}'')\) largest singular values of \(Y^{(i+1)}\), and let \(E^{(i+1)}\subset {\mathbb {R}}^{\ell _{i+1}''}\) be the subspace
Similarly, for every \(1\le j\le i\) and for \(\ell _j\) given by (18), we define the \(x\times \ell _j\) matrix
(again, conditioned on \(\sigma (B_{[r]\times [r+x]}')\), \(Y^{(j)}\) has mutually independent standard normal entries). Denote by \(e^{(j)}_q\), \(1\le q\le \min (\lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor ,\ell _j)\), a random orthonormal system of right singular vectors of \(Y^{(j)}\) corresponding to \(\min (\lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor ,\ell _j)\) largest singular values of \(Y^{(j)}\), and let \(E^{(j)}\subset {\mathbb {R}}^{\ell _{j}}\) be the subspace
Consider the random \(x\times (\ell _{i+1}'+x)\) matrix
Let \({\hat{e}}_1,{\hat{e}}_2,\dots ,{\hat{e}}_{x-\lfloor \tilde{\varepsilon }x\rfloor }\) be a random orthonormal set of right singular vectors of \({\hat{Y}}\) corresponding to \(x-\lfloor \tilde{\varepsilon }x\rfloor \) largest singular values of \({\hat{Y}}\), and let \({\tilde{E}}\subset {\mathbb {R}}^{\ell _{i+1}'+x}\) be the random subspace of dimension \(x-\lfloor \tilde{\varepsilon }x\rfloor \) defined as
Now, we construct the (random) subspace \(H\subset {\mathbb {R}}^{r+x}\) as
Let us check that the constructed subspace satisfies the required lower bound on dimension, that is, \(\dim H\ge (1-(1+\tilde{\varepsilon })^{-i-1})(r+x)\). In view of the assumptions on \(\ell _{i+1}'\), we have
Defining a good event Denote by \(\tilde{\mathcal {E}}\) the event
where the constant \(c'\) is taken from Proposition 3.2. According to our definition of the subspace \({\tilde{E}}\), for every unit vector v as above we have
where the matrix \({\hat{Y}}\) is \(x\times (x+\ell _{i+1}')\) standard Gaussian, in view of the independence of \(\begin{bmatrix} WV^{-1}&Q \end{bmatrix}\) from the \(\sigma \)–field \(\sigma (B_{[r]\times [r+x]}')\).
Hence, by Proposition 3.2 applied to \({\hat{Y}}\), we get
Further, let
and for every \(1\le j\le i\), let
Since, by our assumptions, \(\sqrt{\tilde{\varepsilon }x}/h\ge 2\sqrt{\ell _{i+1}''}\), we have, according to Proposition 3.1,
for a universal constant \(c>0\). Similarly, since for every \(j=1,2,\dots ,i\), \(\sqrt{2^{i+1-j}\tilde{\varepsilon }x}/h\ge \sqrt{\tilde{\varepsilon }x}/h\ge 2\sqrt{x}\), we have
We define
In view of the above,
for a universal constant \({\tilde{C}}>0\).
Checking that H satisfies the required property conditioned on \({\mathcal {E}}_{\textrm{good}}\) Assuming the conditioning, pick any unit vector \(v\in H\). We represent v in terms of the basis \(X_1,\dots ,X_{r+x}\) as
for some coefficients \(a_1,\dots ,a_{r+x}\) with \(\sum _{q=1}^{r+x}a_q^2=1\). Note that
If the last expression is greater than \(\beta ^2\) then we are done. Otherwise, we have
and hence, in particular,
and for every \(j=1,2,\dots ,i\),
Observe that the last conditions yield
In view of conditioning on \(\tilde{\mathcal {E}}\), this immediately implies
Further, in view of conditioning on events \({\mathcal {E}}^{(1)},\dots ,{\mathcal {E}}^{(i+1)}\),
and for every \(j=1,2,\dots ,i\),
Thus, by the triangle inequality,
where the last relation follows from our assumptions on parameters (19) and (13). The assumption on \(\beta \) then implies the result. \(\square \)
Lemma 3.5
There are universal constants \(c\in (0,1]\), \({\tilde{C}}\ge 1\) with the following property. Assume that \(i \in {\mathbb {N}}\), \(\tilde{\varepsilon }\in (0,1]\), \(h\in (0,1]\), r, and x satisfy the assumptions of Proposition 3.3, and assume additionally that
where \(\ell _{i+1}'\) is given in (16). Then with probability at least
we have
where \(\tau \) is defined by (15).
Proof
Construction of subspace H Consider a refinement of the block representation of \(B'\):
where
In particular, for every \(1\le j\le i\), the matrix \(D_j\) above is \(\ell _j\times \ell _j\), \(M_j\) is \(\ell _j\times x\), and \(W_j\) is \(x\times \ell _j\), where \(\ell _j\)’s are given by (18). Further, \(D_{i+1}'\) is \(\ell _{i+1}'\times \ell _{i+1}'\), \(M_{i+1}'\) is \(\ell _{i+1}'\times x\), and \(W_{i+1}'\) is \(x\times \ell _{i+1}'\); the dimensions of \(D_{i+1}''\), \(M_{i+1}''\) and \(W_{i+1}''\) are defined accordingly. In this proof, we denote by \(P_{i+1}':{\mathbb {R}}^{r+x}\rightarrow {\mathbb {R}}^{\ell _{i+1}'}\) the coordinate projection onto first \(\ell _{i+1}'\) coordinates, by \(P_x:{\mathbb {R}}^{r+x}\rightarrow {\mathbb {R}}^{x}\) the coordinate projection onto last x coordinates, and, for every \(1\le j\le i\), denote by \(P_j:{\mathbb {R}}^{r+x}\rightarrow {\mathbb {R}}^{\ell _j}\) the coordinate projection onto \(\ell _j\) components starting from \(1+\sum _{d=j+1}^{i+1}\ell _d\).
Denote by \(e^{(i+1)'}_q\), \(1\le q\le \ell _{i+1}'-\lfloor \tilde{\varepsilon }x\rfloor \), a random orthonormal system of right singular vectors of \(W_{i+1}'\) corresponding to \(\ell _{i+1}'-\lfloor \tilde{\varepsilon }x\rfloor \) largest singular values of \(W_{i+1}'\), and let \({\tilde{E}}\subset {\mathbb {R}}^{\ell _{i+1}'}\) be the subspace
For every \(1\le j\le i\), denote by \(e^{(j)}_q\), \(1\le q\le \min (\lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor ,\ell _j)\), a random orthonormal system of right singular vectors of \(W_j\) corresponding to \(\min (\lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor ,\ell _j)\) largest singular values of \(W_j\), and let \(E^{(j)}\subset {\mathbb {R}}^{\ell _{j}}\) be the subspace
Finally, we construct a random subspace \({\hat{E}}\subset {\mathbb {R}}^x\) as follows. Denote by \(e^{(Q)}_q\), \(1\le q\le \lfloor \tilde{\varepsilon }x\rfloor \), a random orthonormal system of right singular vectors of Q corresponding to \(\lfloor \tilde{\varepsilon }x\rfloor \) largest singular values of Q, and let \(E^{(Q)}\subset {\mathbb {R}}^x\) be the subspace
Further, let \(e^{(M_{i+1}')}_q\), \(1\le q\le \ell _{i+1}'-\lfloor \tilde{\varepsilon }x\rfloor \), be a random orthonormal system of right singular vectors of \(M_{i+1}'\) corresponding to \(\ell _{i+1}'-\lfloor \tilde{\varepsilon }x\rfloor \) largest singular values of \(M_{i+1}'\), and let
For every \(1\le j\le i\), let \(e^{(M_{j})}_q\), \(1\le q\le \lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor \), be a random orthonormal system of right singular vectors of \(M_j\) corresponding to \(\lfloor 2^{j-i-1}\tilde{\varepsilon }x\rfloor \) largest singular values of \(M_{j}\), and let
We then set
The subspace H is now defined as
Let us check that H satisfies the required assumptions on the dimension. We have
Next, we use the assumption on \(\ell _{i+1}'\) and the assumptions on parameters to obtain
Defining a good event Denote by \(\tilde{\mathcal {E}}\) the event
where the constant \(c'\) is taken from Proposition 3.2. According to our definition of the subspace \({\tilde{E}}\), for every unit vector v as above we have
Hence, by Proposition 3.2 applied to \(W_{i+1}'\), we get
Further, for every \(1\le j\le i\), let
Note that conditioned on \({\mathcal {E}}^{(j)}\), we have
Since for every \(j=1,2,\dots ,i\), \(\sqrt{2^{i+1-j}\tilde{\varepsilon }x}/h\ge \sqrt{\tilde{\varepsilon }x}/h\ge 2\sqrt{x}\), we have
Finally, we define events corresponding to a “good” realization of \({\hat{E}}\). Let \({\mathcal {E}}_{M_{i+1}'}\) be the event
Repeating the argument for \(\tilde{\mathcal {E}}\), we get
Similarly, adjusting the argument for \({\mathcal {E}}^{(j)}\) accordingly, we get that for every \(1\le j\le i\), the event
has probability at least \(1-2\exp \big (-c\,2^{i+1-j}\ell _j x\,\tilde{\varepsilon }/h^2\big )\), and that the event
has probability at least
We define
In view of the above,
for a universal constant \({\tilde{C}}\ge 1\).
Checking that H satisfies the required property conditioned on \({\mathcal {E}}_{\textrm{good}}\) Assuming the conditioning, pick any unit vector \(v\in H\). First, we observe that
whereas, by the definition of \({\hat{E}}\) and the conditioning,
Thus, if \(\Vert P_x v\Vert _2\ge \frac{2\sqrt{x}\tau }{c' \lfloor \tilde{\varepsilon }x\rfloor \,h}\) then \(\Vert B'v\Vert _2\ge \Vert M_{i+1}'P_x v\Vert _2-\Vert D_{i+1}'P_{i+1}'v\Vert _2\ge \tau \), and we are done.
Otherwise, if
then, in view of the conditioning (see the definition of \({\mathcal {E}}_{M_{j}}\)),
On the other hand, by our assumptions
Thus, unless \(\Vert B'_{[r]\times [r+x]}v\Vert _2\ge \tau \), we must have
implying
As a final step of the proof, we will show that for any unit vector \(v\in H\) satisfying conditions (22) and (23), one has \(\Vert B'_{\{r+1,\dots ,r+x\}\times [r+x]}v\Vert _2\ge \tau \). First, note that (22) and (23) imply that
whence, in view of conditioning on \(\tilde{\mathcal {E}}\),
Now, for every \(1\le j\le i\), by the above and in view of conditioning on \({\mathcal {E}}^{(j)}\),
whence
Similarly, in view of conditioning on \({\mathcal {E}}_Q\), we get
Thus,
and the proof is complete. \(\square \)
4 Random polytopes, and distances to pivot rows
Let B be an \(n \times m\) matrix with \(m\ge n\), and assume that every square submatrix of B is invertible. We define recursively the sequence of indices \(\{i_r(B)\}_{r \in [n]}\), vectors \(\{v_r(B)\}_{r \in [n]}\) in \({\mathbb {R}}^m\), and polytopes \(\{K_r(B)\}_{r \in [n]}\) in \({\mathbb {R}}^m\) as follows.
Set \(v_1(B):=e_1\) and \(I_0(B):=\emptyset \). For r from 1 to n,
Observe that \(v_r(B)\) is a null vector of \(B_{I_{r-1}, [r]}\) such that \(v_{r}(B)=1\), and that \(i_r(B)\) can be viewed as the index of the r-th pivot row in the Gaussian Elimination with Partial Pivoting with the [rectangular] input matrix B. Note also that our definition of the sets \(I_r(B)\) is consistent with that of the sets \(I_r(A)\) discussed earlier. The above construction does not provide any tie-breaking rules for the choice of the indices \(i_r(B)\) in case when respective expressions have multiple maximizers. In our setting, however (when B is Gaussian), each pivot is unique with probability one, and hence the choice of a tie-breaking rule is irrelevant. We have an immediate relation
where \(\sigma _k\) is the standard Gaussian measure for the corresponding dimension.
Suppose we have performed r steps of the GEPP algorithm on the \(n\times n\) Gaussian matrix A. Let \(I\subset [n]\) have size r, and condition on a realization of \(I_r=I\) and \(A_{I,[r]}\), which determines \(K_r(A)\). Then, for every \(j\in [n]{\setminus } I\), the jth row of A is a Gaussian vector conditioned to stay within the polytope \(K_r(A)\). Formally, for every \(I\subset [n]\) of size r, every \(j\in [n]{\setminus } I\), and every Borel subset \({\mathcal {B}}\) of \({\mathbb {R}}^n\),
We will not directly use the above description of the conditional distribution of \((A_{j,[n]})^\top \) given \(A_{I,[r]}\); instead, we will apply a simple decoupling based on Lemma 4.1 which essentially establishes the same property. We provided the above formula only to clarify our argument.
Lemma 4.1
Suppose B is an \(n\times m\) random matrix such that its entries are i.i.d and have continuous distribution. Then, for \(r \in [n]\) and \(I \subset [n]\) with \(|I|=r\), almost surely the following assertions are equivalent:
-
1.
\(I_r(B)=I\),
-
2.
\(\forall s \in [r],\, i_s(B) = i_s(B_{I,[m]}),\, v_s(B)= v_s(B_{I,[m]}), \hbox { and } K_s(B) = K_s(B_{I,[m]}) \),
-
3.
\(\forall j \in [n]\backslash I\), \((B_{j,[m]})^\top \in K_r(B_{I,[m]})\).
Proof
We can assume without loss of generality that everywhere on the probability space, all square submatrices of B are invertible, and for all \(1\le s\le n-1\) and \(I'\subset [n]\) of size s, the expression
attains its maximum on \(i\in [n]\backslash I'\) at a unique point. These conditions ensure that the above algorithm for generating \(\{i_r(B)\}_{r \in [n]}\), \(\{v_r(B)\}_{r \in [n]}\), \(\{K_r(B)\}_{r \in [n]}\) have a uniquely determined output i.e no ambiguity in the choice of the indices \(i_r(B)\) occurs.
Notice that the implication \(2 \Rightarrow 3\) is straightforward by the above definitions. We will check the implications \(1\Rightarrow 2\) and \(3\Rightarrow 1\) below.
Implication \(1 \Rightarrow 2\). Condition on the event \(\{I_r(B)=I\}\). We have \(v_1(B)=v_1(B_I) = e_1\) and
Further, assume that \(k<r\) is such that \(\forall s \in [k]\), \(i_s(B)= i_s(B_{I,[m]})\). Since \(I_k(B)=I_k(B_{I,[m]})\), we also have \(v_{k+1}(B)=v_{k+1}(B_{I,[m]})\), and thus,
Thus, by induction, \(i_s(B)= i_s(B_{I,[m]})\) for all \(s\in [r]\), whence \(v_s(B)=v_s(B_{I,[m]})\), \(K_s(B)=K_s(B_{I,[m]})\), and \(I_s(B)=I_s(B_{I,[m]})\) for all \(s \in [r]\).
Implication \(3 \Rightarrow 1\). The argument is based on induction just as above. We assume that \(\forall j \in [n]\backslash I\), \((B_{j,[m]})^\top \in K_r(B_{I,[m]})\). First, \(v_1(B)=v_1(B_{I,[m]})=e_1\), and since
we have
On the other hand, by the definition of \(i_1(B_{I,[m]})\),
As a consequence, \(i_1(B)=i_1(B_{I,[m]})\in I\), completing the base step of the induction. Now, let \(k<r\) be an integer such that \(\forall s \in [k], i_s(B)=i_s(B_{I,[m]})\). Since \(v_{k+1}(B)=v_{k+1}(B_{I,[m]})\) by our construction, and since
we get
which implies that \(i_{k+1}(B)=i_{k+1}(B_{I,[m]})\in I\). Thus, we conclude by induction that \(i_s(B)=i_s(B_{I,[m]})\in I\) for all \(s\in [r]\), and the result follows. \(\square \)
As the first main result of the section, we have a probability estimate for the event that the Gaussian measure of the polytope \(K_r(A)\) is below a given threshold:
Proposition 4.2
(Gaussian measure of \(K_r(A)\)) Let A be an \(n\times n\) Gaussian matrix. Then for any \(r \in [n-1]\) and any \(t \ge 2\),
Proof
We start by writing
For each summand, we apply Lemma 4.1 to get
Since \(K_r(A_{I,[n]}) \) and \((A_{i,[n]})^\top \) for \(i\in [n]\backslash I\) are independent, we get
Finally, in view of the standard bound \({ n \atopwithdelims ()n-r} \le n^{n-r} \) for the number of subsets \(I \subset [n]\) of size r, and by the union bound argument, the result follows. \(\square \)
Lemma 4.3
Let A be an \(n\times n\) Gaussian matrix, and let \(r\in [n]\) and \(\tau \in (0,1)\) be parameters. Then, conditioned on the event \(\big \{\sigma _n(K_r(A)) \ge \tau \big \}\),
where H is the subspace of \({\mathbb {R}}^r\) spanned by vectors \((A_{i_s,[r]})^\top \), \(s\in [r-1]\).
Proof
Let \(v:= v_r(A)/ \Vert v_r(A)\Vert _2\) and let \(P:{\mathbb {R}}^n\rightarrow {\mathbb {R}}^r\) be the orthogonal projection onto the span of \(\{e_s\}_{s\in [r]}\). From the definition of v, we have that Pv is a unit normal to the hyperplane H in \({\mathbb {R}}^r\). Then
It remains to note that, by the definition of \(K_r(A)\), on the event \(\big \{\sigma _n(K_r(A)) \ge \tau \big \}\) we have
\(\square \)
As a corollary, we obtain the following probabilistic bound on the distance between \((A_{i_r,[r]})^\top \) and the span of “previous” rows (selected at previous steps of the GEPP process) \((A_{i_s,[r]})^\top \), \(s\in [r-1]\):
Corollary 4.4
Let A be an \(n\times n\) Gaussian matrix. For \( t \ge 2\) and \( r \in [n-1]\), with probability at least \(1-n^{ - t(n-r)/2 }\) we have
where H is the random subspace of \({\mathbb {R}}^r\) spanned by vectors \((A_{i_s,[r]})^\top \), \(s\in [r-1]\).
Proof
In view of Lemma 4.3, the statement would follow as long as \({{\mathbb {P}}}\big \{\sigma _n(K_r(A)) \ge n^{-t}\big \}\ge 1-n^{ - t(n-r)/2 }\). The latter is verified in Proposition 4.2. \(\square \)
5 A recursive argument
The goal of this section is to bound from below the intermediate singular values \(s_{r-k}(A_{I_r,[r]})\) for every r greater than some absolute constant and for k of a constant order. We will start with bounding the intermediate singular values in the bulk of the singular spectrum first and then will recursively apply Proposition 3.3 to provide lower bounds for smaller and smaller intermediate singular values.
As we mentioned in the overview of the proof, the intermediate singular values \(s_{r-k}(A_{I_r},[r])\) for \(k \gg r^{1/2}\,\textrm{polylog}(n)\) can be easily estimated from below with high probability by taking the union bound over the estimates of \(s_{r-k}(A_{I,[r]})\) (see Proposition 3.2) for \(I \subset [n]\) with \(|I|=r\). To bound \(s_{r-k}(A_{I_r,[r]})\) from below for smaller values of k we apply the following strategy. We choose an appropriate positive integer \(r'<r\), condition on a realization of \(I_{r'}\) and \(A_{I_{r'},[r']}\), and, for any I with \(I_{r'} \subset I \subset [n]\) and \(|I|=r\), apply Proposition 3.3 with \(B:=A_{I,[r]}\) and \(F:=A_{I_{r'},[r']}\). This way, \(s_{r-k}(A_{I,[r]})\) is bounded below with high probability conditioned on an event that the intermediate singular values \(s_{r-k'}(A_{I_{r'},[r']})\) are well bounded for every \(k'\) slightly bigger than k.
Definition 5.1
For an integer \(k \in [n]\) and parameters \(p,\beta \ge 1\), let \({\mathcal {E}}_{\textrm{is}}(p,k,\beta )\) be the event that
and \({\mathcal {E}}_{\textrm{rec}}(p,k,\beta )\) be the event that
Note that although n is not mentioned explicitly in the list of parameters for \({\mathcal {E}}_{\textrm{is}}(p,k,\beta )\), it clearly depends on the underlying matrix dimension.
The next proposition is the main result of this section:
Proposition 5.2
There is a universal constant \(C>0\) with the following property. Let \(p \ge 1\). Then there exist positive integers \(n_0:=n_0(p)\), \(120p\le k_0:=k_0(p)\le Cp\), and a positive real number \(300p \le \beta _0:=\beta _0(p)\le Cp\), so that for any \(n\ge n_0\) and \(\beta \ge \beta _0\),
We remark that the lower bounds on \(k_0(p)\) and \(\beta _0(p)\) in the assumptions of the proposition are not required in the proof but will be needed later. As a corollary of the proposition (proved in the end of this section), we have
Corollary 5.3
For any \(p\ge 1\), \(\beta \ge \beta _0(p)\),
Now, we present a technical version of the above proposition. We introduce several “section-level” parameters. Let \(\tilde{\varepsilon }>0\) be a small constant and L be a large integer to be determined later. The parameter \(\tilde{\varepsilon }\) will play the same role as in Proposition 3.3. Next, let
and let \(s_1\) be the smallest integer such that \( 2^{s_1} m_0 \ge n\). Then we define the finite sequence \(m_1,\dots ,m_{s_1+1}\), where
The main technical result in this section is the following
Lemma 5.4
Fix \(\tilde{\varepsilon }\in (0, 1/100]\) and \(L\ge 1/ \tilde{\varepsilon }\). Then there exists a positive integer \(n_0\) (depending on \(\tilde{\varepsilon }\) and L) such that for any \(n \ge n_0\) and \(s \in [0, s_1-1]\), we have for every \(\alpha \ge 4\):
where \( c( \tilde{\varepsilon })\) and \(C( \tilde{\varepsilon })\) are positive constants which depend on \(c, {\tilde{C}}\) from Proposition 3.3 and on \( \tilde{\varepsilon }\).
Proof of Proposition 5.2
Let \(\tilde{\varepsilon }:= 1/100\). We can safely assume that the constants \( c( \tilde{\varepsilon })\) and \(C( \tilde{\varepsilon })\) from Lemma 5.4 satisfy \(c( \tilde{\varepsilon })\in (0,1]\) and \(C( \tilde{\varepsilon })\ge 1\). Choose
Let \(\beta _0:= \max \{ 4 c(\tilde{\varepsilon })L,300p\}\), \(k_0(p):= \max \{ \lceil 9\,L/ \tilde{\varepsilon }\rceil , 120p\}\), and let \(\beta \ge \beta _0\). Applying Lemma 5.4 with \(\alpha \ge 4\) satisfying \( \beta /(40p) = C(\tilde{\varepsilon })\alpha \), we get
implying the result for large enough n. \(\square \)
For the rest of the section, we fix \(s \in [0, s_1-1]\).
5.1 Choice of parameters and the growth function
Definition 5.5
(Definition of \(i_{th},i_{\max }\), \(f_i\), \(r_i\)) For a given positive integer L and for \( \tilde{\varepsilon }\in (0,1/4]\), let \(i_{th}\) be the integer such that
and let \(i_{\max }\) be the integer such that
Note that \(\tilde{\varepsilon }m_s/10\ge L/\tilde{\varepsilon }\), and hence \(i_{th}\le i_{\max }\).
For every \(i \in [i_{th},i_{\max }]\), we define a non-decreasing function
Further, we define a collection of integers \(\{r_i\}_{i\in [i_{\max }+1]}\) inductively as follows. Whenever \(i \in [i_{th}]\), we set \(r_i:= m_s \). Further, assuming that \(r_i\) has been defined for some \(i\in [i_{th},i_{\max }]\), we let \(r_{i+1}\) be the smallest integer such that \(f_i(r_{i+1}) \ge r_i\). Note that \(m_s=r_1\le r_2 \le \dots \le r_{i_{\max }+1}\).
We recall our strategy: to bound the singular value \(s_{\lfloor ( 1 - ( 1+ \tilde{\varepsilon })^{-i-1})r \rfloor }(A_{I_r,[r]})\) from below, we will select an appropriate integer \(r'<r\) and apply Proposition 3.3 with \(B:= A_{I,[r]}\) and \(F:=A_{I_{r'},[r']}\), taking the union bound over all subsets \(I\subset [n]\) with \(I_{r'}\subset I\) and \(|I|=r\). The function \(f_i\) defined above, determines the choice of \(r'\), namely, we choose
for \(r_i \le r \le m_{s+2}\). The indices \(i_{th}\) and \(i_{\max }\) defined above, determine the range of application for the inductive strategy; namely, \(i_{\max }\) marks the largest index i for which our induction argument can be applied, and \(i_{th}\) indicates a threshold value below which the corresponding singular values \(s_{\lfloor ( 1 - ( 1+ \tilde{\varepsilon })^{-i})r \rfloor }(A_{I_r,[r]})\) concentrate very strongly and can bounded directly with help of Proposition 3.2 and a simple union bound argument.
The goal of this subsection is to verify certain relations between the introduced parameters, that need to be satisfied in order to apply the results on the singular values established earlier. Since the results here are of purely computational nature, we present the proofs in the Appendix.
Lemma 5.6
(Inequalities for \(i_{\max }\)) Let \( \tilde{\varepsilon }\in (0, 1/4)\) and \(L\ge 1/ \tilde{\varepsilon }\). For \(r \in [m_s, m_{s+2} ]\),
Further,
Lemma 5.7
(Assumptions in Proposition 3.3) Let \(\tilde{\varepsilon }\in (0, \frac{1}{28})\) and \(L\ge 4\). Fix \(i \in [i_{th}, i_{\max }]\) and assume that \({\tilde{r}}\) satisfies \(r_{i+1}\le {\tilde{r}}\le m_{s+2}\). Let \( r: = f_i( {\tilde{r}}) \) and \( x:= {\tilde{r}} - r\). Then,
Moreover, i, r, x, and \(\tilde{\varepsilon }\) satisfy the assumptions in Proposition 3.3, specifically,
For a given \(i \in [i_{th}, i_{\max }]\), the number \({\tilde{r}}\) satisfying the assumptions of the above lemma can only be chosen if \(r_{i+1}\le m_{s+2}\). In the next statement, we show that the inequality is satisfied for every admissible i (and in fact verify a slightly stronger bound):
Lemma 5.8
(An upper bound on \(r_{i_{\max }+1}\)) Let \( \tilde{\varepsilon }\in (0, 1/28)\) and \(L\ge 4\). Then, \(r_{i_{\max }+1}\le 2m_s=m_{s+1}\).
To construct the growth function \(g(\cdot )\) from (13), we first define an auxiliary positive function \(g_{s}(\cdot )\), and then set
for all admissible j. The formal definition of \(g_s(\cdot )\) is given below.
Definition 5.9
Let \(\alpha \ge 1\) be a parameter. For \( i \in [i_{th}]\), we set
where \(c'\) is the constant from Proposition 3.2.
For \(i \in [i_{th},i_{\max }]\), we apply a recursive definition:
where \(h_s(i)\) is given by
and where \(C_h\ge -\log \big (2^{-11}(c')^2\tilde{\varepsilon }\big )\) is a constant depending only \(c, {\tilde{C}}\) (from Proposition 3.3) and \(c'\), and which we shall determine in Lemma 5.10.
The function \(h_s(i)\) corresponds to the parameter h in Proposition 3.3, and is constructed in such a way that certain union bound argument that we are going to apply further works. The next lemma clarifies the choice of the constant \(C_h\) from the above definition:
Lemma 5.10
The constant \(C_h\) can be chosen so that the following holds. For \(i \in [i_{th}, i_{\max }]\) and \( {\tilde{r}} \in [r_{i+1}, m_{s+2}]\), let \(r:= f_i( {\tilde{r}} )\) and \( x:= {\tilde{r}} - r\). Then,
In the next lemma we verify the crucial bound on the growth function which will ultimately guarantee a polynomial in n bound on the intermediate singular values:
Lemma 5.11
There exists \(C( \tilde{\varepsilon })>1\) which depends on \(c', {\tilde{C}}\) from Proposition 3.3, on \(C_h\), and on \( \tilde{\varepsilon }\), such that
5.2 Good events, and probability estimates
Definition 5.12
For \( i \in [i_{\max }+1]\) and \(r \in [r_i, n]\), let \( {\mathcal {E}}(r, i)\) be the event that
where \(g_s(\cdot )\) is given in Definition 5.9 and \(r_i,i_{\max }\) are taken from Definition 5.5. Further, we denote \( {\mathcal {E}}(r, [i]):= \bigcap _{ j \in [ i ] } {\mathcal {E}}(r,j)\).
Lemma 5.13
For \(\tilde{\varepsilon }\in (0, 1/100]\), \(L\ge 4\) and \( \alpha \ge 4\),
Proof
Fix \(i \in [i_{th}]\) and \(r \in [m_s, m_{s+2}]\). Let \(q:= r - \lfloor (1 - (1+ \tilde{\varepsilon })^{-i})r \rfloor \). Then
We recall that in view of the definition of \(m_s\) and \(i_{th}\), necessarily \(q<r\); furthermore,
For each \(I \subset [n]\) with \(|I|=r\),
Applying (40), we conclude that
We complete the proof with the union bound argument. There are \({n \atopwithdelims ()r} \le (en/r)^r \le \exp ( r\log n) \) subsets \(I \subset [n]\) with \(|I|=r\). As \(r \le m_{s+2}\le 4m_s\), and in view of the definition of \(m_s\) and our choice of \(\tilde{\varepsilon }\),
By applying the union bound argument again over all \(i \in [i_{th}]\) and all \( r \in [m_s, m_{s+2}]\), the statement of the lemma follows. \(\square \)
Lemma 5.14
Assume \(\tilde{\varepsilon }\in (0, 1/100]\), \(L\ge 4\) and \( \alpha \ge 4\). For \( i \in [i_{th},i_{\max }]\) and \( {\tilde{r}} \in [r_{i+1}, m_{s+2}] \), set \(r:= f_i( {\tilde{r}})\) and \(x:={\tilde{r}}-r\). Then
where the random polytope \(K_r(A) \subset {\mathbb {R}}^n\) was defined in (24), and where \(\sigma _n\) is the standard Gaussian measure in \({\mathbb {R}}^n\).
Proof
We start by noting that the last inequality in the statement of the lemma follows from the estimate \(x \ge (1 + \tilde{\varepsilon })^{-i_{\max }}m_s \ge L/ \tilde{\varepsilon }\) (see Lemma 5.7 and the definition of \(i_{\max }\)).
We further partition the event in question so that
For each \(I \subset [n]\) with \(|I|=r\), we define \({\mathcal {E}}(I)\) to be the event
and note that for each admissible I, \({\mathcal {E}}(r, [i])\cap \{ I_r(A) = I\} \subset {\mathcal {E}}(I)\).
For \(I\subset [n]\) with \(|I|=r\) and \(J \subset [n] \backslash I \) with \(|J| = x \), let \({\mathcal {E}}(I, J)\) be the event that
Denote \(K(I):= K_r(A_{I,[r]})\subset {\mathbb {R}}^r\). Then each term in (41) can be bounded as
Now, assume that for every \(I \subset [n]\) and \(J \subset [n] \backslash I\) with \(|I| = r\) and \(|J|=x\),
Clearly, for each \(I \subset [n]\) with \(|I|=r\),
Together with (42) and (43), this gives
and the result follows.
Thus, it remains to show (43). By Lemma 4.1, almost everywhere on the probability space we have
Hence,
In view of the joint independence of the entries of A, we obtain
where the outer expectation is with respect to \(A_{I,[r]}\).
For each realization of \(A_{I,[r]}\) such that the event \({\mathcal {E}}(I)\) holds, we apply Proposition 3.3 with
to bound \( {{\mathbb {P}}}( {\mathcal {E}}(I,J)^c\, |\,A_{I,[r]} )\). Let \(g(\cdot )\) be a growth function satisfying
where \(g_s(\cdot )\) is given in Definition 5.9. Since \(16\, g_s(j+1) \le g_s(j)\) for \(j \in [i-1]\) and \(g_s(j) \le 1\) for \( j \in [i]\), the function \(g(\cdot )\) defined this way satisfies (13). Recall that on the event \({\mathcal {E}}(I)\) we have
We apply Proposition 3.3 with g(t) and with \(h:=h_s(i)\) (see Definition 5.9) so that
(observe that our parameters r, x satisfy the assumption of the proposition due to Lemma 5.7, and that h satisfies the assumption \(h\le 2^{-11} (c')^2 \tilde{\varepsilon }\) in view of the assumptions on the constant \(C_h\) in Definition 5.9). We get
In view of Lemma 5.10, this implies
Combining the last inequality with (45), we obtain
Next, we will treat the denominator in the estimate (44). By Fubini’s theorem,
Almost everywhere on the event \(\{\sigma _r(K(I)) \ge n^{-\alpha /2}\}\) we have
whence
Together with (46) and (44), this yields
and the proof of (43) is complete. \(\square \)
At this point, we are ready to prove the main lemma in this section.
Proof of Lemma 5.4
First, recall that in view of Lemma 5.8, \(r_{i_{\max }+1}\le m_{s+1}\), and that in view of (31) we have \(r - \lceil 9/ \tilde{\varepsilon }\rceil \le \lfloor (1-(1+\tilde{\varepsilon })^{-i_{\max }+1})r\rfloor \), whence
where we used the definition of the events \({\mathcal {E}}(r,i)\) (Definition 5.12). To estimate the probability of the union of the events in the last line, we shall embed it into a specially structured collection.
Let \(r':= f_{i_{\max }}(m_{s+2})\), where \(f_{\cdot }(\cdot )\) was defined in (30). We have
To be able to apply a recursive bound from the last lemma, we use the bounds \(r_i\le f_i( {\tilde{r}} ) \le m_{s+2}\), \( {\tilde{r}} \in [r_{i+1}, m_{s+2}]\), \(i \in [i_{th}+1,i_{\max }]\), to write
Thus, using that \(f_i( {\tilde{r}} )\le r'\) for \(i \in [i_{th}+1,i_{\max }]\) and that \(m_{s+2} - r' \ge L \), we get
for some \(c( \tilde{\varepsilon })>0\). It remains to note that in view of Lemma 5.11, \(g_s(i_{\max }+1)\ge n^{-C(\tilde{\varepsilon })\alpha }\) for some \(C(\tilde{\varepsilon })\). \(\square \)
Proof of Corollary 5.3
For brevity, we denote \(k_0:=k_0(p)\). We fix \(r \in [k_0+1,n-2k_0]\) and let \(F_{r} \subset {\mathbb {R}}^{I_r}\) be the right singular subspace of the matrix \((A_{I_r,[r]})^\top \) corresponding to \(k_0\) smallest singular values of \((A_{I_r,[r]})^\top \) (since almost everywhere on the probability space \(I_r\) is unambiguously determined, and the singular values of \((A_{I,[r]})^\top \) are distinct for every \(I\subset [n]\) with \(|I|=r\), \(F_r\) is uniquely defined). Now, let us define the event \(\tilde{\mathcal {E}}_r(\beta )\) that
where \((A_{I_r,[r+1,r+2k_0]})^\top \big \vert _{F_r} \) and \((A_{I_r,[r+1,r+2k_0]})^\top \big \vert _{F_r^\perp }\) are linear operators obtained by restricting the domain of \((A_{I_r,[r+1,2k_0]})^\top \) to \(F_r\) and \(F_r^\perp \), respectively. Then, conditioned on the intersection \(\tilde{\mathcal {E}}_r(\beta ) \cap {\mathcal {E}}_{\textrm{is}}(p,k_0,\beta )\), for any \(v \in {\mathbb {R}}^{I_r} \backslash \{0\}\),
where \(P_{F_r}\) and \(P_{F_r^\perp }\) are orthogonal projections onto \(F_r\) and \(F_r^\perp \), respectively. Consider two cases.
-
Suppose \( \Vert P_{F_r^\perp }v\Vert _2 \ge \frac{1}{4} \frac{n^{-\beta /(40p)}}{ 3\sqrt{\beta n } } \Vert P_{F_r}v\Vert _2\). Then,
$$\begin{aligned} \Vert v\Vert _2&= \sqrt{ \Vert P_{F_r^\perp }v\Vert _2^2 + \Vert P_{F_r} v\Vert _2^2 } \le \Vert P_{F_r^\perp }v\Vert _2 \sqrt{1^2 + \Big ( \frac{1}{4} \frac{n^{-\beta /(40p)}}{ 3\sqrt{\beta n } }\Big )^{-2} } \\&\le O(\sqrt{\beta })\,n^{\beta /(40p) + \frac{1}{2}}\,\Vert P_{F_r^\perp }v\Vert _2, \end{aligned}$$which implies
$$\begin{aligned} n^{-\beta /(50p)} \Vert P_{F_r^\perp }v\Vert _2 \ge O(\beta ^{-1/2})\,n^{-\beta / (50p) - \beta /(40p) - \frac{1}{2}}\,\Vert P_{F_r^\perp }v\Vert _2 \ge n^{-\beta /(20p)} \Vert v\Vert _2, \end{aligned}$$where the last inequality holds because \(\beta \ge \beta _0(p) \ge 300p\) and since n is sufficiently large depending on p.
-
In the case \( \Vert P_{F_r^\perp }v\Vert _2 < \frac{1}{4} \frac{n^{-\beta /(40p)}}{ 3\sqrt{ \beta n } } \Vert P_{F_r}v\Vert _2\), we have
$$\begin{aligned} n^{-\beta /(40p)} \big \Vert P_{F_r}v\Vert _2 - 3\sqrt{\beta n} \big \Vert P_{F_r^\perp }v\big \Vert _2 \ge \frac{3}{4} n^{-\beta /(40p)} \Vert P_{F_r}v\Vert _2 \ge n^{-\beta /(20p)} \Vert v\Vert _2. \end{aligned}$$
Since the above estimate holds for all \(v \in {\mathbb {R}}^{I_r}\backslash \{0\}\), we conclude that everywhere on the intersection \(\tilde{\mathcal {E}}_r(\beta ) \cap {\mathcal {E}}_{\textrm{is}}(p,k_0,\beta )\),
Therefore, for \(p\ge 1\) and \(\beta \ge \beta _0(p)\),
and thus
Since in view of Proposition 5.2, \({{\mathbb {P}}}\big ( {\mathcal {E}}_{\textrm{is}}(p,k_0,\beta )^c \big ) \le n^{-2\beta +o_n(1)}\), the corollary will follow if we show that \({{\mathbb {P}}}( \tilde{\mathcal {E}}_r(\beta )^c )\le n^{-2\beta -1 +o_n(1)}\).
From now on, we fix \(r \in [k_0+1, n-2k_0]\) and condition on a realization of \(A_{[n],[r]}\) such that the set \(I_r\) and the space \(F_r\) are uniquely determined. We will write \(\tilde{{\mathbb {P}}}\) and \(\tilde{\mathbb {E}}\) to denote the corresponding conditional probability and conditional expectation.
The independence of the entries of the matrix A implies that \(Q:=(A_{I_r,[r+1,r+2k_0]})^\top \big \vert _{F_r}\) and \(W:=(A_{I_r,[r+1,r+2k_0]})^\top \big \vert _{F_r^\perp }\) are (standard) Gaussian linear operators from \(F_r\) to \({\mathbb {R}}^{2k_0}\) and from \(F_r^\perp \) to \({\mathbb {R}}^{2k_0}\), respectively. For the purpose of estimating the operator norm and least singular values, we can view W and Q as matrices with i.i.d N(0, 1) entries of dimensions \(2k_0\times (r-k_0)\) and \(2k_0\times k_0\), respectively; more specifically, we can define standard Gaussian matrices \({\tilde{W}}\) and \({\tilde{Q}}\) of dimensions \(2k_0\times (r-k_0)\) and \(2k_0\times k_0\) such that everywhere on the probability space the singular spectrum of W and \({\tilde{W}}\), and of Q and \({\tilde{Q}}\), agree.
It is well known that the expected operator norm of any standard Gaussian matrix is no more than the sum of square roots of its dimensions (see, for example, [20, Section 7.3]). Hence,
Since the spectral norm is 1-Lipschitz, the standard Gaussian concentration inequality (see, for example, [20, Section 5.2]) implies
Next, we derive an estimate for \(s_{\min }(Q)=s_{\min }({\tilde{Q}})\). For \(i \in [k_0]\), let \(P_i:{\mathbb {R}}^{k_0}\rightarrow {\mathbb {R}}^{k_0}\) be the orthogonal projection to the subspace which is orthogonal to the columns vectors \({\tilde{Q}}_{[2k_0],j}\) for \(j \in [k_0]\backslash \{i\}\). Then,
Since \(P_j\) and \({\tilde{Q}}_{[2k_0],j}\) are independent, the norm \(\Vert P_j({\tilde{Q}}_{[2k_0],j})\Vert _2\) is equidistributed with that of a \(2k_0-(k_0-1)=(k_0+1)\)–dimensional standard Gaussian vector. Since the probability density function of a \((k_0+1)\)–dimensional Gaussian vector is bounded above by \((2\pi )^{-(k_0+1)/2}\), we obtain
where \(|B_2^{k_0+1}|\) is the Lebesgue measure of the unit Euclidean ball \(B_2^{k_0+1}\) in \({\mathbb {R}}^{k_0+1}\). Therefore, in view of the previous computations,
Applying the bound \((|B_2^{k_0+1}|)^{1/(k_0+1)} = O( k_0^{-1/2})\), we get that there exists a universal constant \(C_{\textrm{b}}\ge 1\) so that
Now, setting \(t:= n^{-\beta /(40p)}\), we get
where the last inequality holds since \(k_0=k_0(p)\ge 120p\).
As a final step of the proof, rewriting (49) and (50) on the entire probability space, we get
We conclude that \( {{\mathbb {P}}}( {\mathcal {E}}_r(\beta )^c ) \le n^{-2\beta -1 +o_n(1)}\), and the result follows. \(\square \)
6 The smallest singular value and the growth factor in exact arithmetic
6.1 Distance to subspaces
Recall that by \(i_t=i_t(A)\), \(1\le t\le n\), we denote the indices of the pivot rows in the GEPP process (see Sect. 4).
Definition 6.1
(Subspaces generated by row vectors of submatrices of A) For \(x,r \in [n]\) with \(1\le x \le r\), let
and let \(H_{r,0}:=\{0\}\). Additionally, for \(s \in [x]\), let
where we set \(H_{r,1,1}:=\{0\}\).
Definition 6.2
For \(\beta >0\), let \({\mathcal {E}}_{\textrm{row}}(r,\beta )\) be the event that
and set
Further, let \({\mathcal {E}}_{\textrm{dist}}(\beta )\) be the event that
The goal in this section is to prove
Proposition 6.3
There exists \(\beta _{6.3} \ge 2\) so that the following holds. For \(\beta \ge \beta _{6.3}\), we have \({\mathcal {E}}_{\textrm{dist}}(\beta ) \supset {\mathcal {E}}_{\textrm{row}}(\beta )\), and
The statement is obtained as a combination of Lemmas 6.6 and 6.7 below. First, we consider two simple facts from Euclidean geometry.
Lemma 6.4
Let \(u \in {\mathbb {R}}^r\) and let \(H \subset {\mathbb {R}}^r\) be a subspace. Then for any orthogonal projection P in \({\mathbb {R}}^r\), we have
Proof
The statement follows immediately by observing that P is a contraction. \(\square \)
Lemma 6.5
Let F be a subspace of \({\mathbb {R}}^k\), and let \(v_1, v_2 \in {\mathbb {R}}^k\) be vectors such that
For \(i \in [2]\), let \(F_i\) be the linear span of F and \(v_i\). Then,
Proof
For any subspace E, we let \(P_E\) be the orthogonal projection onto E. Let \(u_i:= \frac{P_{F^\perp }v_i }{ \Vert P_{F^\perp }v_i \Vert _2}\). Observe that,
whence
On the other hand,
and therefore
\(\square \)
Lemma 6.6
Let \(s,k,r \in [n-1]\) such that \(s\le r-k < r\). Fix a realization of A such that the event \({\mathcal {E}}_{\textrm{row}}(\beta )\) holds. Then,
Thus, \({\mathcal {E}}_{\textrm{dist}}(\beta ) \supset {\mathcal {E}}_{\textrm{row}}(\beta )\).
Proof
First, we note that for every \(t \in [2,r]\),
where \(P_t: {\mathbb {R}}^r \mapsto {\mathbb {R}}^t\) is the coordinate projection onto the first t components. Applying Lemma 6.4 for every \(2\le t\le r\), we obtain
where in the last inequality we used the definition of \({\mathcal {E}}_{\textrm{row}}(\beta )\). Further, for \(t \in [r-k+1,r]\), we will apply Lemma 6.5 with
so that \(F_1 = H_{r,t-1}\) and \(F_2 = {H}_{r,t,s}\), and from the previous inequality and the definition of \({\mathcal {E}}_{\textrm{row}}(\beta )\) we have
Lemma 6.5 implies
(it is easy to see that in the case \(\dim \mathrm{span\,}\{F,v_1,v_2\}<\dim (F)+2\) when the lemma cannot be applied, the above inequality holds as well). Together with the inequality \( \sqrt{n} + 3 \sqrt{ \beta \log n } \le 2n^{1+ \beta / (n-t) }\) for \(\beta >0\),
Finally, applying the above inequality inductively for t from \(r-k+1\) to r we obtain
\(\square \)
Lemma 6.7
For \(\beta \ge 2\) and \(1\le r\le n\), the following probability estimate holds:
and
Proof
First, in view of Corollary 4.4, we have
Next, for each \(i \in [n]\), applying the standard concentration inequality for Lipschitz functions of Gaussian variables,
With \({\mathbb {E}}\Vert A_{i,[n]}\Vert _2 \le ({\mathbb {E}}\Vert A_{i,[n]}\Vert _2^2)^{1/2}\le \sqrt{n}\), by taking \(t:= 3\sqrt{ \beta \log n }\) we have
Taking the union bound over \(i\in [n]\) and taking into account the condition \(\beta \ge 2\), we get the first assertion of the lemma.
The second assertion follows from another application of the union bound. \(\square \)
6.2 The smallest singular value of \(A_{I_r,[r]}\)
Definition 6.8
For \(k \in [n]\), \(\beta ,p \ge 1\), let \({\mathcal {E}}_{\textrm{sq}}(p,k,\beta )\) be the event that
Proposition 6.9
There is a universal constant \(C>0\) with the following property. For any \(p \ge 1\), there exist \(n_0(p)\), \(k_1(p)\le Cp^2\) and \( \beta _6.3\le \beta _1(p)\le Cp^2\) such that for \( n \ge n_0(p), \beta \ge \beta _1(p)\), and \(k_1(p)\) we have
where \(k_0(p)\) is taken from Proposition 5.2, \(\beta _6.3\) is defined in Proposition 6.3, and \({\mathcal {E}}_{\textrm{rec}}(\cdot )\) is taken from Definition 5.1. Moreover,
Proof
Note that if the events’ inclusion above holds then the second assertion of the proposition follows immediately by combining the bounds \({{\mathbb {P}}}( {\mathcal {E}}_{\textrm{rec}}(p,k_0(p),\beta )^c ) \le n^{-2\beta +o_n(1)}\) from Corollary 5.3 and \({{\mathbb {P}}}( {\mathcal {E}}_\textrm{row}(\beta )^c)\le n^{-\beta }\) from Lemma 6.7. Thus, we can focus on proving the first assertion.
Let \(k_1(p)=\beta _1(p) \ge 400 k_0(p)\,p\) where \(k_0(p)\) is taken from Proposition 5.2.
Consider an argument by contradiction. Fix any realization of A such that both \({\mathcal {E}}_{\textrm{row}}(\beta )\) and \({\mathcal {E}}_{\textrm{rec}}(p, k_0(p),\beta )\) hold, and such that for some \(r \in [k_1(p), n-k_1(p)]\), \(s_{r}(A_{I_r,[r]}) \le n^{-\beta /(6p)}\), that is, there exists a unit vector \(u \in {\mathbb {R}}^{I_r}\) such that
(we assume here that the column of the matrix \((A_{I_r,[r]})^\top \) are indexed over the set \(I_r\)). Since \(\Vert u\Vert _2=1\), there is an index \(s \in [r]\) such that \( |u_{i_s}|\ge r^{-1/2} \ge n^{-1/2} \), whence
Thus, our realization of A and our choice of s satisfy
Set \(k:= \min \{ 2\,k_0(p), r-s\}\).
Assume first that \(s<r\), i.e \(k>0\). In view of the inclusion \( {\mathcal {E}}_{\textrm{row}}(\beta ) \subset {\mathcal {E}}_{\textrm{dist}}(\beta )\) (see Proposition 6.3), we get
Since \(n-r \ge k_1(p)\) and \(\beta \ge \beta _1(p)= k_1(p) \ge 400 k_0(p)\,p \ge 200kp\), we have
whence
Further, in the situation when \(k=0\), the inequality (53) is still true as can be immediately seen from (52).
Next, we will show that (53) leads to contradiction. The argument depends on whether \(k=r-s\) or not.
Case 1 \(k=r-s\). By the definition of the event \({\mathcal {E}}_{\textrm{row}}(\beta )\), we have
where we used that \(n-s \ge n-r \ge k_1(p)\ge 200k_0(p)\,p\). In view of the condition \(\beta \ge 200k_0(p)\,p\ge 400p\),
which contradicts (53).
Case 2 \(k=2k_0(p)<r-s\). In this case, \((A_{i_s,[r]})^\top \) is a column vector of \((A_{I_{r-k}, [r]})^\top \) and \(H_{r,r-k,s}\) is the span of every other column vector \((A_{i_{s'},[r]})^\top \) for \(s' \in [r-k]\backslash \{s\}\). Hence, in view of (53),
However, this contradicts the definition of the event \({\mathcal {E}}_\textrm{sq}(p, k_0(p),\beta )\):
The result follows. \(\square \)
The next simple lemma will be used to show that with high probability indices of the pivot rows obtained in exact arithmetic coincide with the results of the floating point computations.
Lemma 6.10
There is a universal constant \(C>0\) and a number \(n_0\in {\mathbb {N}}\) such that, assuming \(n\ge n_0\),
Proof
In view of Proposition 6.9 (say, applied with \(p=1\)), there are constants \(C_1,n_0>0\) such that, assuming \(n\ge n_0\),
For indices \(r<k_1(1)\), we use the trivial union bound:
where in the last line we used the standard bound on the smallest singular value of a square Gaussian random matrix [4, 17]. Similarly, we get
Combining the three estimates above, we get the result. \(\square \)
6.3 Estimating the growth factor in exact arithmetic
Definition 6.11
For \(\beta >1\), let \({\mathcal {E}}_{\textrm{col}}(\beta )\) be the event that
and for \(\tau >1\), let \({\mathcal {E}}_{\textrm{entry}}(\tau )\) be the event that
Lemma 6.12
For any \(\beta \ge 2\), we have
furthermore, for every \(\tau \ge 1\),
Proof
The upper bound on \({{\mathbb {P}}}( {\mathcal {E}}_{\textrm{col}}(\beta )^c)\) can be derived exactly the same way as in the argument for \({\mathcal {E}}_\textrm{row}(\beta )\) (see the proof of Lemma 6.7), so we skip the discussion.
To estimate the complement of \({\mathcal {E}}_{\textrm{entry}}(\tau )\), we write
where in the last inequality we used that the probability density function of the standard Gaussian random variable is bounded by \(\frac{1}{\sqrt{2\pi }}\). \(\square \)
At this point, we are ready to prove the “exact arithmetic” counterpart of the main statement of the paper:
Proposition 6.13
There is a universal constant \(C>1\) and a function \(\tilde{n}:[1,\infty )\rightarrow {\mathbb {N}}\) with the following property. Let \(p\ge 1\), and let \(n\ge {\tilde{n}}(p)\). Then
Proof
Recall that the parameter \(k_1(p)=O(p^2)\) was defined in Proposition 6.9. We can take a universal constant \(C_1>0\) large enough so that \(C_1p^3 \ge 600 p\,k_1(p)\) for all \(p \ge 1\). Fix \(\beta \ge C_1p^3\), set \(\tau :=\beta /(100p)\), and assume \(n\ge \sqrt{100p}\). In view of the assertions of Lemma 6.7, Proposition 6.9, and Lemma 6.12, in order to show that
(which would imply the statement), it is sufficient to verify that everywhere on the intersection
we have
In what follows, we use the notation introduced at the beginning of Sect. 2; in particular, we work with matrices \({\mathcal {M}}^{(\ell )}\), \(0\le \ell \le n-1\), defined in (3).
Recall that
For \(r \in [k_1(p)]\), we simply use the bound above to get
Further, for \(r \in (k_1(p),n-k_1(p)]\), we write
In view of formula (4) and our conditioning on the event \({\mathcal {E}}_{\textrm{sq}}(p,k_1(p),\beta )\), for \(s\in [k_1(p),n-k_1(p)]\), \(i \in [n]\backslash I_s\) and \(j >s\), we have
and thus
and for every \(s\in [r]\) with \(s>k_1(p)\),
By our earlier observation,
Combining the estimates together, we conclude that for all \(r \in [n-k_1(p)]\),
For the “last” \(k_1(p)\) admissible values of r, we rely on (56) again to get
In the end, we make use of our bound \(\beta /(6p) \ge 100 k_1(p)\) to conclude that for all large n,
This completes the proof. \(\square \)
7 GEPP in floating point arithmetic
In this section we transfer the statement of Proposition 6.13 into the proper context of the floating point arithmetic. We expect a part of the argument in this section (specifically, in the proof of Lemma 7.2) to be rather standard for experts in numerical analysis. Still, we prefer to provide all the details to make the paper self-contained.
Lemma 7.1
Let A be an \(n\times n\) Gaussian matrix and \(A={\mathcal {M}}^{(0)},{\mathcal {M}}^{(1)},\dots , {\mathcal {M}}^{(n-1)}\) be the sequence of matrices generated by the GEPP in exact arithmetic (see (3)). Then, for every \(1\le k\le n-1\),
Proof
Fix any \(1\le k\in n-1\). With the vector \(v_{k}(A)\) defined at the beginning of Sect. 4 and in view of (4), for every \(i \notin I_{k-1}(A)\) we have
Fix any subset \(I\subset [n]\) of cardinality \(k-1\) and any \((k-1) \times k\) matrix M, and condition on the realizations \(I_{k-1}(A)=I\) and \(A_{I_{k-1},[k]}=M\). In what follows, we denote the conditional probability measure by \(\tilde{{\mathbb {P}}}\). Under this conditioning, \(v_{k}(A)\) and the polytope \(K:=K_{k-1}(A)\) (see Sect. 4; here we adopt the convention \(K_0(A):={\mathbb {R}}^n\)) are fixed. For \(i \notin I\), let
By Lemma 4.1, under the conditioning the vectors \(A_{i,[n]}\) for \(i\notin I\) are i.i.d., with the probability density function
which is symmetric and log-concave (i.e \( y \mapsto \log (\rho (y))\) is a concave function). Since log-concavity is preserved under taking marginals, the random variable \( \langle \frac{v_{k}(A)}{\Vert v_{k}(A)\Vert _2},(A_{i,[n]})^{\top }\rangle \) is also log-concave and symmetric under the conditioning. This implies, in particular, that the probability density function \(\rho _X\) of \(X_i\)’s (\(i \notin I\)) is non-increasing on the positive semi-axis.
Now, since \(X_{i_{k}} = \max _{i \notin I}X_i\), we have
whereas
Combining the two identities and using the monotonicity of \(\rho _X\), we get
The result follows by applying Fubini’s theorem. \(\square \)
Lemma 7.2
Let M an \(n\times n\) invertible matrix and \({\hat{M}}:= \textrm{fl}(M)\). Let \(PM=LU\) be the PLU-factorization of M in exact arithmetic, assume that \(P=\textrm{Id}_n\), and let \(M=M^{(0)},M^{(1)},\dots ,\) \(M^{(n-1)}=U\) be the sequence of matrices obtained during the elimination process. Let \(\delta \in (\textbf{u},1/3)\) be a parameter and assume that the matrix M and the unit roundoff \(\textbf{u}\) satisfy
and
Then GEPP in floating point arithmetic succeeds for \({\hat{M}}\); the computed permutation matrix \({\hat{P}}=\textrm{Id}_n\), and, denoting by \({\hat{M}} = {\hat{M}}^{(0)}, {\hat{M}}^{(1)},\dots , {\hat{M}}^{(n-1)}\) the sequence of matrices obtained during the elimination process, for every \(k=0, 1,\dots ,n-1\),
Proof
We will prove the statement by induction. Fix any \(k\in [n-1]\). Assume that all of the following holds
-
(a)
The computed matrix \({\hat{M}}^{(k-1)}\) has been produced by taking indices of the first \(k-1\) pivot rows to be \(1,2,\dots ,k-1\), and \( |{\hat{M}}^{(k-1)}_{k,k}| > \max _{i \in [k+1,n] }|{\hat{M}}^{(k-1)}_{i,k}|\), so that the index of the k-th computed pivot row is k.
-
(b)
\({\hat{M}}^{(k-1)} = G_{k-1}\cdots G_1(M+{\tilde{E}}^{(k-1)})\), where \(G_i\) is the Gauss transformation to eliminate i-th row of \({\hat{M}}^{(i-1)}\), \(1\le i\le k-1\), and where the error matrix \({\tilde{E}}^{(k-1)}\) satisfies
$$\begin{aligned} \Vert {\tilde{E}}^{(k-1)} \Vert \le 8kn\,\textbf{u}\, \max _{i,j,\ell }|M_{i,j}^{(\ell )}|. \end{aligned}$$ -
(c)
\(\max \limits _{i,j}|{\hat{M}}^{(v)}_{i,j}|\le 2\,\max \limits _{i,j,\ell }|M^{(\ell )}_{i,j}|\) for all \(0\le v\le k-1\).
Note that, by the assumptions on the matrix M, the induction hypothesis for the base case \(k-1=0\) is satisfied.
Let \(G_k=\textrm{Id}_n-\tilde{\tau }^{(k)}\,e_k^\top \) be the Gauss transformation which eliminates entries \({\hat{M}}^{(k-1)}_{i,k}\), \(i=k+1,\dots ,n\), so that in exact arithmetic we have
The computed matrix \({\hat{M}}^{(k)}\) can be explicitly written as
(note that we “force” \({\hat{M}}^{(k)}_{i,k}\) to be 0 for \(i\in [k+1,n]\) whereas the f.p. expression \(\textrm{fl}\big (\textrm{fl}\big (\frac{ {\hat{M}}^{(k-1)}_{i,k}}{ {\hat{M}}^{(k-1)}_{k,k}}\big )\, {\hat{M}}^{(k-1)}_{k,k}\big )\) is not necessarily equal to \({\hat{M}}^{(k-1)}_{i,k}\)). Denote
Since the first k rows of \(E^{(k)}\) are 0, for every \(i \in [k]\) we have \(G_i E^{(k)}=E^{(k)}\), so \({\hat{M}}^{(k)}\) can be expressed in the form
Applying the above equality together with the induction hypothesis, we obtain
where \({\tilde{E}}^{(k)}:={\tilde{E}}^{(k-1)}+E^{(k)}\). Note that non-zero entries of \(E^{(k)}\) are all contained within the bottom right \((n-k)\times (n-k)\) submatrix of \(E^{(k)}\), and for every \(i,j \in [k+1,n]\) we have
since there are 3 floating point operations and \( \frac{ |{\hat{M}}^{(k-1)}_{i,k}|}{ |{\hat{M}}^{(k-1)}_{k,k}|}\le 1\), and where in the last inequality we used the induction assumption (c). Thus,
confirming condition (b) on the kth step. Moreover, in view of the assumptions on M we then have
Further, by the assumptions on \(s_{\min }(M_{[k],[k]})\) and in view of the bound on the norm of \(E^{(k)}\), the matrix \((M+\tilde{E}^{(k)})_{[k],[k]}\) is invertible. Hence, for every \(i\in [k+1,n]\) there is a unique linear combination \(L_i\) of the first k rows of \(M+{\tilde{E}}^{(k)}\) such that the vector \(\textrm{row}_i(M+\tilde{E}^{(k)})-L_i\) has first k components equal zero. We conclude that necessarily the matrices \(G_1,G_2,\dots ,G_k\) are Gauss transformations for \(M+{\tilde{E}}^{(k)}\), whence for every \(j \in [k+1,n]\) we have
whereas
We will rely on formulas (60) and (61) to show that \({\hat{M}}^{(k)}\) and \(M^{(k)}\) are sufficiently close entry-wise.
In view of (60) and (61), for every \(j\in [k+1,n]\) we have
Note that
and that the bound \(2\Vert {\tilde{E}}^{(k)}\Vert \le s_{\min }(M_{[k],[k]})\) implies
Thus, applying (59), for every \(i=k+1,\dots ,n\) we get
Immediately we have
By the nature of the Gaussian Elimination process, the first k rows of \({\hat{M}}^{(k)}\) coincide with those of \({\hat{M}}^{(k-1)}\), and the bottom left \((n-k)\times k\) submatrix of \({\hat{M}}^{(k)}\) is zero. And thus
confirming the condition (c) for the kth step.
It remains to check the condition (a). Note that we only need to consider the case \(k\le n-2\). Using the definition of vectors \(v_k(\cdot )\) from the beginning of Sect. 4, we can write
where \(H \subset {\mathbb {R}}^{k+1}\) is the subspace spanned by the first k rows of \(M_{[k+1],[k+1]}\). Applying (62) and (57), we conclude that
and the result follows. \(\square \)
Proof of Theorem A
In view of Lemma 6.10, there are \(C',n_0>0\) such that, assuming \(n\ge n_0\),
On the other hand, standard concentration estimates for the spectral norm of Gaussian matrices (see, for example, [20, Chapter 4]) implies that, assuming n is sufficiently large,
Let \(\delta \in (\textbf{u},1/3)\) be a parameter to be chosen later, and define the events
The above observations on the spectral norm and the smallest singular values imply
Further, Proposition 6.13 (applied, say, with \(p=2\)) yields for all sufficiently large n,
for some universal constant \(C'''>0\). Finally, in view of Lemma 7.1, we have
Thus, the intersection of the events \({\mathcal {E}}_1(\delta )\cap {\mathcal {E}}_2(\delta )\cap {\mathcal {E}}_3(\delta )\) has probability at least
Taking \(\delta :=\textbf{u}^{1/8}\), we get that for any large enough n,
where \({\tilde{C}}>0\) is a universal constant. It remains to note that, in view of Lemma 7.2, everywhere on the intersection \({\mathcal {E}}_1(\textbf{u}^{1/8})\cap {\mathcal {E}}_2(\textbf{u}^{1/8}) \cap {\mathcal {E}}_3(\textbf{u}^{1/8})\) the GEPP in floating point arithmetic succeeds for \({\hat{A}}^{(0)}=\textrm{fl}(A)\); the computed permutation matrix \({\hat{P}}\) coincides with the matrix P from the PLU–factorization of A in exact arithmetic, and
A second application of Proposition 6.13, now to bound \(\textbf{g}_{\textrm{GEPP}}(A)\) conditioned on the intersection \({\mathcal {E}}_1(\textbf{u}^{1/8})\cap {\mathcal {E}}_2(\textbf{u}^{1/8}) \cap {\mathcal {E}}_3(\textbf{u}^{1/8})\), completes the proof. \(\square \)
8 Further questions
In this section, we mention some open questions related to the probabilistic analysis of the Gaussian Elimination with Partial Pivoting.
Sharp estimate of the growth factor Our main result shows that with probability close to one, the growth factor of GEPP is at most polynomial in the matrix dimension, \(\textbf{g}_{\textrm{GEPP}}(A) \le n^C\). Our analysis leaves the constant \(C>0\) unspecified, and it would be of interest to obtain an estimate with a reasonable (single digit) explicit constant. Furthermore, as we mentioned in the introduction, it was suggested in [5] based on numerical simulations that for large n, \(\textbf{g}_{\textrm{GEPP}}(A) =O(n^{1/2+o_n(1)})\) with probability close to one. The problem of finding the optimal constant power of n in the growth factor estimate seems to require essential new ideas. At the same time, it is natural to expect that recurrent estimates of the singular spectrum of submatrices obtained in the GEPP process should remain a key element of the future refinements of our result.
The probability that the Gaussian Elimination with Partial Pivoting succeeds in the floating point arithmetic Our main result states that, under the assumption that the dimension n is sufficiently large, with probability at least \(1-\textbf{u}^{1/8}\,n^{{\tilde{C}}}\) the GEPP in f.p. arithmetic succeeds for \(\textrm{fl}(A)\), and the computed permutation matrix agrees with that obtained in exact arithmetic. We expect the probability estimate to be much stronger, perhaps of the form \(1-\textbf{u}^{1-o_n(1)}\,n^{{\tilde{C}}}\), and leave this as an open problem.
Smoothed analysis of the growth factor Our proof does not extend to the smoothed analysis setting without incurring significant losses in the upper estimate for the growth factor. In fact, our treatment of the partially random block matrices B in Sect. 3 heavily relies on the assumption that the norm of a submatrix within the “random part” of B is typically of the same order as the square root of that submatrix’ larger dimension. Establishing a polynomial upper bound on the growth factor in the presence of a non-random shift (of polynomial operator norm) is an interesting and challenging problem.
Notes
It can be checked, for example, that with a constant probability
$$\begin{aligned} \max \limits _{J\subset [n],\,|J|=n/2;\,j\in [n]{\setminus } J} \big \Vert \textrm{row}_j(A_{[n],[n/2]}) \big (A_{J,[n/2]} \big )^{-1}\big \Vert _2 \end{aligned}$$is greater than any predefined constant power of n. We expect that a much stronger lower bound can be established.
References
Boucheron, S., Lugosi, G., Massart, P.: Concentration Inequalities. Oxford University Press, Oxford (2013)
Bourgain, J., Tzafriri, L.: Invertibility of “large’’ submatrices with applications to the geometry of Banach spaces and harmonic analysis. Israel J. Math. 57(2), 137–224 (1987)
Chafaï, D.: Singular values of random matrices (2009)
Edelman, A.: Eigenvalues and condition numbers of random matrices. SIAM J. Matrix Anal. Appl. 9(4), 543–560 (1988)
Edelman, A., Mascarenhas, W.: On the complete pivoting conjecture for a Hadamard matrix of order \(12\). Linear Multilinear Algebra 38(3), 181–187 (1995)
Golub, G.H., Van Loan, C.F.: Matrix Computations. Johns Hopkins Studies in the Mathematical Sciences, 4th edn. Johns Hopkins University Press, Baltimore (2013)
Higham, N.J.: Accuracy and Stability of Numerical Algorithms, 2nd edn. Society for Industrial and Applied Mathematics (SIAM), Philadelphia (2002)
Higham, N.J.: How accurate is Gaussian Elimination? Technical report (1989)
Litvak, A.E., Tikhomirov, K., Tomczak-Jaegermann, N.: Small ball probability for the condition number of random matrices. In: Geometric Aspects of Functional Analysis. Vol. II, pp. 125–137, Lecture Notes in Math., 2266. Springer, Cham
Livshyts, G., Paouris, G., Pivovarov, P.: On sharp bounds for marginal densities of product measures. Israel J. Math. 216(2), 877–889 (2016)
Naor, A., Youssef, P.: Restricted invertibility revisited. In: A Journey Through Discrete Mathematics, pp. 657–691. Springer, Cham (2017)
Nguyen, H.H.: Random matrices: overcrowding estimates for the spectrum. J. Funct. Anal. 275(8), 2197–2224 (2018)
Rudelson, M., Vershynin, R.: Small ball probabilities for linear images of high-dimensional distributions. Int. Math. Res. Not. IMRN 2015(19), 9594–9617 (2015)
Sankar, A.: Smoothed Analysis of Gaussian Elimination. Ph.D. thesis. MIT (2004)
Sankar, A., Spielman, D.A., Teng, S.-H.: Smoothed analysis of the condition numbers and growth factors of matrices. SIAM J. Matrix Anal. Appl. 28(2), 446–476 (2006)
Spielman, D.A., Srivastava, N.: An elementary proof of the restricted invertibility theorem. Israel J. Math. 190, 83–91 (2012)
Szarek, S.J.: Condition numbers of random matrices. J. Complex. 7(2), 131–149 (1991)
Trefethen, L.N., Schreiber, R.S.: Average-case stability of Gaussian elimination. SIAM J. Matrix Anal. Appl. 11(3), 335–360 (1990)
Vershynin, R.: John’s decompositions: selecting a large part. Israel J. Math. 122, 253–277 (2001)
Vershynin, R.: High-Dimensional Probability. Cambridge Series in Statistical and Probabilistic Mathematics, vol. 47. Cambridge University Press, Cambridge (2018)
Wilkinson, J.H.: Error analysis of direct methods of matrix inversion. J. Assoc. Comput. Mach. 8, 281–330 (1961)
Yeung, M.-C., Chan, T.F.: Probabilistic analysis of Gaussian elimination without pivoting. SIAM J. Matrix Anal. Appl. 18(2), 499–517 (1997)
Funding
Open Access funding provided by Carnegie Mellon University.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
K.T. was partially supported by the Sloan Research Fellowship and by the NSF Grant DMS 2054666.
Appendices
Proof of Proposition 3.2
In this section, we provide a proof of Proposition 3.2 which is based on the restricted invertibility argument of Nguyen [12].
The classical restricted invertibility theorem was derived by Bourgain–Tzafriri [2]; see, in particular, [11, 16, 19] for extensions and a comprehensive discussion. In what follows, we will use the estimate from paper [16]:
Theorem A.1
[16] Let B be an \(u\times t\) (\(t\ge u\)) matrix. Then for any \(\varepsilon \in (0,1)\) with \(\frac{\varepsilon ^2 \Vert B\Vert _{HS}^2}{\Vert B\Vert ^2}\ge 1\) there exists a subset \(J\subset [t]\) such that
Further, we recall an anti-concentration estimate for orthogonal projections of random vectors with independent components of bounded density, obtained in [13] (see also [10] for improved constants).
Theorem A.2
[13] Let X be a random vector in \({\mathbb {R}}^\ell \) with independent components having the distribution densities uniformly bounded by \(\rho >0\). Then for every \(d\le \ell \) and every d-dimensional subspace \(H\subset {\mathbb {R}}^\ell \), the distribution density of the orthogonal projection of X onto H is bounded above by \((C\rho )^d\), where \(C>0\) is a universal constant.
Now, we are ready to prove Proposition 3.2. Although the proof essentially repeats the one in [12], we provide the full argument for completeness. The argument works for matrices M with i.i.d entries with continuous distribution density uniformly bounded above by \(\rho >0\).
Proof of Proposition 3.2
In view of the interlacing properties of singular values (see, for example, [3, Theorem 1.4]), we can assume without loss of generality that \(t=u\). We fix an index \(1\le i\le u-1\), and \(s\in (0,1]\). Denote by \({\mathcal {E}}\) the event
where \(c=c(\rho )>0\) will be chosen later. Let \(Z^\top \) be the \(u\times i\) random orthogonal matrix measurable w.r.t \(\sigma (M)\) and such that \(\Vert M\textrm{col}_j(Z^\top )\Vert _2\le \frac{ci\,s}{\sqrt{u}}\), \(1\le j\le i\), everywhere on \({\mathcal {E}}\) (one may take \(Z^\top \) as the matrix whose columns are the normalized right singular vectors of M corresponding to i smallest singular values of M).
In view of Theorem A.1, there is a (random) subset \(J\subset [u]\) such that
everywhere on the probability space.
In particular, since \(s_{|J|} (Z_{[i]\times J})>0\), there exists an \(i\times |J|\) matrix Y such that \((Z_{[i]\times J})^\top Y=I_i\) and hence \(\Vert Y\Vert \le \big (1-2^{-1/2} \big )^{-1} \sqrt{u/i}\) everywhere.
Define a subspace
and let P be the orthogonal projection onto \(H_J^\perp \). We then have
whence on the event \({\mathcal {E}}\),
So, we can write
To estimate the probabilities on the right hand side of the inequality, we apply Theorem A.2. Observe that, in view of Theorem A.2,
for some universal constant \(C'>0\). Thus,
The result follows by choosing a sufficiently small \(c=c(\rho )\). \(\square \)
Proofs of the auxiliary results in Sect. 5.1
Proof of Lemma 5.6
As a verification of (31), we have
For the upper bound of \(i_{\max }\), we have
which implies
where we used that \( \log (1+t) \ge \frac{t}{2}\) for \(0\le t \le \frac{1}{4}\). \(\square \)
Proof of Lemma 5.7
First, from the definition of \(f_i\) and since \(r=f_i({\tilde{r}})\) we have
Thus, the lower bound for x in (33) is obtained. Further, in view of the definition of \(i_{\max }\), and since \(r\ge r_i\ge m_s\),
Combining this with the last inequality in (63), we obtain
Next, we will verify the conditions on the parameters which appear in Proposition 3.3. For convenience, we itemize the rest of the proof according to the inequalities we wish to check:
-
1.
\(\textbf{r} - \lfloor (\textbf{1} - ( \textbf{1} + \tilde{\varepsilon })^{-\textbf{i}}) \textbf{r} \rfloor \le \textbf{x} \le \textbf{r}.\) First, the inequality \(( 1 + \tilde{\varepsilon })^{-i}r \le x\) together with x is an integer implies \( x \ge \lceil ( 1 + \tilde{\varepsilon })^{-i} r \rceil \). Since
$$\begin{aligned} \forall t \in [0,r],\, \lfloor r-t \rfloor + \lceil t \rceil = r, \end{aligned}$$we have
$$\begin{aligned} r - \lfloor (1 - ( 1 + \tilde{\varepsilon })^{-i}) r \rfloor =\lceil ( 1 + \tilde{\varepsilon })^{-i} r \rceil \le x. \end{aligned}$$For the upper bound on x, by the definition of \(i_{th}\) we have \( ( 1 + \tilde{\varepsilon })^{-i_{th}}m_s \le m_s/8 \), which implies
$$\begin{aligned} x \le \frac{21}{20} ( 1 + \tilde{\varepsilon })^{-i} r \le \frac{21}{20} (1 + \tilde{\varepsilon })^{-i_{th}} \cdot 4m_s \le m_s \le r. \end{aligned}$$ -
2.
\(\tilde{\varepsilon } \textbf{x} \ge 4\) and \(\tilde{\varepsilon } ( 1 + \tilde{\varepsilon } )^{-\textbf{i}}{} \textbf{r} \ge \textbf{2}\). This follows immediately as
$$\begin{aligned} \tilde{\varepsilon }x \ge \tilde{\varepsilon }( 1 + \tilde{\varepsilon })^{-i} r \ge \tilde{\varepsilon }( 1 + \tilde{\varepsilon })^{- i_{\max }} m_s \ge L \ge 4. \end{aligned}$$ -
3.
\(3 ( 1 + \tilde{\varepsilon })^{-\textbf{i}-\textbf{1}} \textbf{r} - ( 1+ \tilde{\varepsilon })^{-\textbf{i}}{} \textbf{r} \ge \textbf{x} + \textbf{1} + \textbf{11} \tilde{\varepsilon }\textbf{x}.\) Since \( \tilde{\varepsilon }\le \frac{1}{28}\), we have, in view of the above,
$$\begin{aligned} 3 ( 1 + \tilde{\varepsilon })^{-i-1} r - ( 1+ \tilde{\varepsilon })^{-i}r \ge \frac{3}{2} ( 1+ \tilde{\varepsilon })^{-i}r \ge \frac{10}{7} x, \end{aligned}$$whereas \(x + 1 + 11 \tilde{\varepsilon }x \le x + 12 \tilde{\varepsilon }x \le \frac{10}{7} x\). \(\square \)
Proof of Lemma 5.8
First, by the definition of \(r_i\)’s we have \(r_i= m_s\) for \( i \in [i_{th}]\). Next, assume that the statement is true for all indices in an interval [i], for some \(i \in [i_{th},i_{\max }]\). We will show it holds for \(i+1\). Fix for a moment any \(i'\in [i_{th},i]\). By (33) (applied with \(x=r_{i'+1}-r_{i'}= r_{i'+1} - f_{i'}(r_{i'+1})\)) we have
The definition of \(i_{th}\) and the conditions on \(\tilde{\varepsilon }\) imply \(\frac{21}{20}(1+ \tilde{\varepsilon })^{-i_{th}}\cdot 2m_s \le \frac{ \tilde{\varepsilon }}{2} m_s\), and we obtain
Finally, summing up over all \(i'\in [i_{th},i]\), we obtain
The result follows. \(\square \)
Proof of Lemma 5.10
For the first summand, from the definition of \(h_s(i)\) and since \(x \ge ( 1 + \tilde{\varepsilon })^{-i} r \ge ( 1 + \tilde{\varepsilon })^{-i}m_s\) (see Lemma 5.7), we get
where the last inequality follows as \(x \le \tilde{r} \le m_{s+2}\le n\). Thus,
As \( \tilde{\varepsilon }x/2 \ge 2 \) (see Lemma 5.7), for \(n \ge 2\) we have
For the second and third summands, since \( ( 1 + \tilde{\varepsilon })^{-i} r \ge \frac{ 20 }{ 21 } x\), we obtain
Now, we impose first constraint on \(h_s(i)\):
which can be achieved by taking \(C_h\) large enough depending only on \({\tilde{C}}\) and c. Then, since \( \tilde{\varepsilon }x \ge 4\),
Next, as the function \(t \mapsto t^2 \log (1/t)\) converges to 0 when \(t \searrow 0\), we can suppose by taking \(C_h\) large enough that
which implies
Finally, combining the last estimate with (65), we obtain
and the lemma follows since \( h_s(i) \le \exp \big ( - \frac{128 \alpha \log n }{ \tilde{\varepsilon }^2 ( 1+ \tilde{\varepsilon })^{-i} m_s } \big ) \) and \( x \ge ( 1 + \tilde{\varepsilon })^{-i}r \ge ( 1 + \tilde{\varepsilon })^{-i} m_s\). \(\square \)
Proof of Lemma 5.11
First, as \(i_{th}\le 2\log (m_s)/ \tilde{\varepsilon }\le 2 \log (n) / \tilde{\varepsilon }\) (see (32)) and in view of the definition of \(g_s(i)\), we have
To estimate \(g_s(i_{\max }+1)\), we write
By the definition of \( h_s(i)\),
Returning to (68), we obtain
where we used \(i_{\max }\le 2 \log (m_s)/ \tilde{\varepsilon }\le 2 \log (n) / \tilde{\varepsilon }\), \( \tilde{\varepsilon }( 1+ \tilde{\varepsilon })^{-i_{\max }}m_s \ge L \ge 1\), and \( \sum _{j=0}^\infty ( 1 + \tilde{\varepsilon })^{-j} = \frac{1+ \tilde{\varepsilon }}{ \tilde{\varepsilon }}\). Together with (67) and since \(\alpha \ge 1\), we can simplify the bound to have the form
\(\square \)
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Huang, H., Tikhomirov, K. Average-case analysis of the Gaussian elimination with partial pivoting. Probab. Theory Relat. Fields 189, 501–567 (2024). https://doi.org/10.1007/s00440-024-01276-2
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00440-024-01276-2