WorksMathematics
Ultra-hard-edge universality under deterministic preconditioning
Abstract
Let be an real matrix with iid entries distributed as a fixed centered, variance-one, symmetric subgaussian random variable. For deterministic invertible , set and . We prove as , uniformly over and . Here , where is uniform on . The constants and depend only on the entry law. No condition-number bound or effective-rank assumption is required. The coefficient tends to one for diffuse inverse profiles. For convergent concentrated or spiked profiles, its limit is an explicit Gaussian expectation. The proof combines the submicroscopic comparison of Sah, Sahasrabudhe, and Sawhney with signed-permutation averaging and a weighted comparison for random hyperplane normals. Applications include inverse amplification, square sample covariance matrices with arbitrary deterministic positive-definite population matrices, and independent random environments.
Introduction and main results
The anisotropic lower tail
For a square iid random matrix, multiplication by a deterministic matrix stretches different directions by different amounts. We ask how this changes the probability that the resulting product is exceptionally close to singular. Our result separates the effect of this anisotropy into an overall scale, determined by the Hilbert–Schmidt norm of the inverse multiplier, and a leading coefficient given by a spherical average. The coefficient records whether the inverse multiplier acts comparably in many directions or concentrates its action in a few, while remaining universal across the entry distributions considered here.
Throughout the paper, is a fixed real random variable satisfying
Let be an matrix whose entries are independent copies of . For a deterministic invertible matrix , set
Here is the squared Hilbert–Schmidt norm. For with and with , define
Thus depends only on the eigenvalues of . For any nonzero deterministic ,
For , one has , , and . The ordinary hard-edge scale is then . We use the term ultra-hard edge for the smaller thresholds considered below, where lies between an exponentially small lower cutoff and an inverse-polynomial upper cutoff. The theorem determines the leading coefficient throughout this shrinking range, uniformly over the preconditioner.
Theorem 1.1 (Deterministic preconditioning). Assume (1). There are constants , , , and a deterministic sequence , depending only on the entry law, such that for every deterministic invertible matrix and every
with defined in (2),
All bounds are uniform over . In particular, throughout (4),
The coefficient is uniformly nondegenerate:
We prove this and more precise profile statements in § 7. Since , the explicit residual in (5) is uniformly negligible in (4). The inequality ensures that the additive exponential remainder is negligible relative to every threshold in that window. The theorem makes no assertion about the optimal exponential endpoint.
Corollary 1.2 (Gaussian-reference universality). Let have iid standard real Gaussian entries. Under the assumptions and in a common window of the form (4),
uniformly over deterministic invertible and admissible .
Proof. Apply Theorem 1.1 to both entry laws, shrink the window to their intersection, and use (7).
Diffuse and concentrated inverse profiles
Define
The quantity tends to zero exactly when the inverse singular mass is diffuse. Theorem 1.1 includes this regime, but does not require it.
Corollary 1.3 (Diffuse coefficient). If , then uniformly in (1.4),
More precisely, (1.5) remains valid with replaced by , provided is added to its right-hand side.
At the other extreme, when the normalized inverse singular mass approaches rank one, the coefficient tends to . More generally, if finitely many weights survive and the residual mass is diffuse, the coefficient tends to
Thus deterministic anisotropy survives through a Gaussian coefficient even though the entry distribution disappears from the leading term.
Related work
Edelman computed the limiting distribution of the least singular value in the Gaussian case [1]. Tao and Vu established universality of this distribution for iid matrices under a sufficiently high finite-moment assumption, with a polynomially small additive error uniform over thresholds [11]. For correlated complex Gaussian Wishart matrices with a fixed difference between the matrix dimensions and population eigenvalues bounded above and away from zero, Hachem, Hardy, and Najim obtained Bessel-kernel limits at the hard edge [3]. These results describe distributions on the ordinary hard-edge scale. Relative asymptotics for exponentially small lower-tail probabilities require error estimates that remain negligible compared with those probabilities. A polynomially small additive error alone is insufficient.
Rudelson and Vershynin established linear lower-tail bounds for square iid subgaussian matrices and the corresponding optimal rectangular estimates [8, 9]. Sah, Sahasrabudhe, and Sawhney (SSS) proved the sharp unpreconditioned coefficient at exponentially small relative thresholds [10]. Letwin and Polavarapu identified as the general preconditioned scale under substantially weaker entry assumptions [5]. Their theorem supplies an upper bound with an unspecified constant. The present result identifies the coefficient under (1.1), uniformly over deterministic invertible preconditioners, without a condition-number bound or an effective-rank assumption. The covariance consequence in Corollary 8.2 concerns real iid latent entries satisfying (1.1) and allows arbitrary positive-definite population matrices, with no uniform spectral bounds, at the ultra-small thresholds specified there.
The inverse-tail mechanism used here originated in the Gaussian phase-retrieval argument at the critical injectivity threshold [4 Section 9.1, Lemma 7]. There, conditioning on the common rows of overlapping square minors leads to a weighted inverse-tail problem for an independent square Gaussian block. Its leading coefficient involves the spherical average , where is Haar-uniform on . For a nonzero $n\times m input matrix , this average satisfies
Thus the same coefficient functional already occurs in that Gaussian argument. The earlier inverse-tail estimate uses the Haar distribution of Gaussian singular vectors and their independence from the singular values, together with moment assumptions on the random input that force its normalized profile to be diffuse. The present paper establishes universality of this coefficient for symmetric iid subgaussian entries, uniformly over arbitrary deterministic input profiles, including concentrated ones. The signed-permutation arguments below replace the exact rotational invariance used in the Gaussian analysis.
Proof strategy
The proof separates the rare-event estimate from the identification of its coefficient. A one-column inverse decomposition produces the mark
where is obtained by deleting one column of and is a unit normal to its column space. This mark measures the inverse multiplier in the random normal direction. It need not concentrate when the inverse profile has a few large eigenvalues.
The first ingredient is the marked transfer theorem, Theorem 5.2. It extends the SSS unmarked regularity and conditional replacement estimates to marks whose second moments are bounded on every signed-permutation orbit. For , averaging over signed permutation matrices gives exactly one. This identity permits truncation at with rare-event cost . On the retained set, the unmarked estimates apply at the enlarged threshold . The resulting coefficient is , where is a signed-permutation-invariant spectral intensity.
The second ingredient is the weighted coefficient comparison, Theorem 3.3. It identifies this expectation uniformly in , using an bound for with and weak convergence of two coordinate magnitudes of . No independence between the normal and the intensity is required. The comparison follows from a transport inequality for signed-permutation averages. The coordinate limit is established in Lemma 3.4 and Appendix § 11 using the projected-column method of Nguyen and Vu [6], with an explicit finite-population variance estimate for the sampling normalization. Combining the two ingredients gives the universal coefficient .
Organization. Section 2 gives the inverse decomposition. Sections 3–5 develop the weighted comparison, the spectral intensity, and the marked transfer theorem. Section 6 combines these ingredients to prove Theorem 1.1. Sections 7 and 8 describe the coefficient and its applications. Sections 9 and 10 discuss the scope of the method and formulate extensions, with exact Gaussian identities. Appendix § 11 proves the coordinate limit.
Conventions
We write for the Euclidean operator norm and use the subgaussian norm
Constants may depend on the fixed entry law. The notation always denotes a deterministic quantity tending to zero. When matrices or thresholds vary, uniformity is stated explicitly. Singular values of a deleted matrix are ordered increasingly as . Square least singular values are denoted by . All assertions concerning a normal vector are insensitive to its global sign. Rank-deficient deleted matrices are excluded from geometric events and assigned intensity zero. Their probability is . In inverse-amplification statements, we count a singular as a failure and assign its inverse norm the value .
The logarithmic cutoffs and regularity events below are defined for , with sufficiently large. Uniform moment bounds for the associated intensities are understood over this range. The value of may be increased in the asymptotic arguments.
The inverse geometry
Write , where and . On choose with orthonormal columns spanning , and choose a measurable unit vector . Put
Then .
Proposition 2.1 (One-column decomposition). For any deterministic , on ,
If
then
Proof. Block inversion proves (2.1). The second term has rank at most one, with norm
Since , the singular values of are the nonzero singular values of . The triangle inequality gives the remaining assertions. Also , so is independent of the choice of .
For subsequent unconditional probability estimates, set on .
For a nonnegative -measurable mark , let
Also put .
Lemma 2.2 (Exact event inclusion). For every , . If is iid as in (1.1), then
Proof. The vector from (2.1) satisfies and . Hence , including when . The probability bound follows from the square least-singular-value estimate [8].
For define on full rank, and zero otherwise. Proposition 2.1 gives, for , outside singularity and ,
The rectangular estimate [9] gives
Signed-permutation averaging
Let denote the finite group of signed permutation matrices of order . All averages over use its uniform probability measure. For fixed ,
Under (1.1), has the same law as , while the normal line transforms as .
Large marks inside an invariant rare event
The following elementary lemma is the marked-regularity input.
Lemma 3.1 (Orbit truncation). Let a finite group act on a random object with invariant law. Let be increasing invariant events satisfying
Suppose satisfies for every . Then, for and ,
If is invariant and integrable, then
Proof. Split into the intervals , . Monotonicity, invariance of , and Markov’s inequality on each orbit give
Summation gives (3.2). For the second assertion, average over the group and use .
For the norm mark, (3.1) gives the exact identity
and consequently . The full events are invariant under simultaneous signed permutations of and . Lemmas 2.2 and 3.1 therefore imply
This is a bound on the marked rare event itself. No exceptional set of merely polynomial probability is discarded.
Transport of orbit-averaged marks
For and , put
Write for the quadratic Wasserstein distance on probability measures on .
Lemma 3.2 (Orbit transport). For all and all rectangular with ,
Moreover .
Proof. Let be the increasingly sorted coordinate magnitudes. They represent the same signed-permutation orbits as . Couple their orbit averages using the same , and put . The reverse triangle inequality, Cauchy–Schwarz, and (18) give
Increasing matching is the optimal quadratic coupling of the two equally weighted one-dimensional empirical measures, so the last expression is . Finally , and Jensen’s inequality gives the upper bound.
Theorem 3.3 (Weighted coefficient comparison). Let be a random unit line and a real random variable whose joint law is invariant under , . Suppose
where are independent standard Gaussians, and suppose for some . Then
No independence between and is assumed.
Proof. Let . For bounded continuous , exchangeability and (24) show that
Applying this to a countable convergence-determining class gives in probability. Both measures have second moment exactly one. The characterization of convergence by weak convergence and convergence of second moments therefore yields
This implication can also be applied along almost surely convergent subsubsequences. Since the measures are supported on and have second moment one, their squared distance is at most two. Thus (26) holds in every fixed , .
For an independent Haar vector , the same assertion follows by writing and applying the law of large numbers. By Lemma 3.2,
Here . Joint invariance gives
Hölder’s inequality with and (27) finish the proof. All bounds are independent of and of its number of columns.
Lemma 3.4 (Two coordinate magnitudes). Let be an matrix with iid centered, variance-one, subgaussian entries. On full rank, let be a unit vector spanning and on rank deficiency extend its line by an independent Haar line. Then
where are independent standard Gaussians.
Proof. Appendix A gives a proof using coordinate delocalization and the projected-column comparison of Nguyen–Vu [6], with an explicit finite-population variance estimate for the sampling normalization. Only weak convergence of two magnitudes is asserted. The exponentially small rank-deficiency probability does not affect this limit.
Under (1.1), the Haar-line extension preserves signed-permutation invariance. Assigning the spectral weight zero on rank deficiency also preserves the joint invariance required by Theorem 3.3.
The regularized spectral intensity
Choose a measurable orthonormal eigenbasis of as a function of that matrix alone, and form the corresponding left singular vectors . On full rank these obey , including with a fixed measurable choice inside repeated eigenspaces. This convention makes all coordinate-flatness and truncated spectral conditions below invariant under .
Put and
For iid , let
We shall use the SSS deleted-matrix event specified in Proposition 5.1. In particular it is invariant and implies . Define
This definition sets off , including on rank deficiency. It is independent of the rare-event threshold and of the preconditioner.
Proposition 4.1 (Subcritical intensity moments). For every ,
The same conclusion holds if is replaced by any event contained in , for fixed .
Proof. It suffices to consider . Jensen’s inequality and subadditivity of the power give
We claim
Let consist of the first columns of . Eigenvalue interlacing for the principal submatrix of gives . By the rectangular estimate [9 Theorem 1.1],
Since , the variable therefore satisfies
On it obeys . Fix such that . For sufficiently large , uniformly for , and tail integration yields
Here , and the last term is uniformly bounded for . This proves (4.6). Summing (4.5) gives . Smaller positive follow by monotonicity of moments. The argument with a general fixed logarithmic exponent is identical. □
Remark 4.2 (Why second moments are unnecessary). The raw Gaussian deficiency-one intensity has a critical second moment: the lower tail of has quadratic order, while the intensity is at least a constant times . The proof uses a fixed and does not require a second-moment estimate for the intensity.
Lemma 4.3 (Gaussian linearization). For full-rank , define
where is independent of . On , for ,
Proof. For every , . On , , and . Conditioning on proves the claim. □
A marked submicroscopic transfer
We first state the precise unmarked facts used from SSS. Keeping the full regularity event separate from the deleted-matrix event is essential.
Proposition 5.1 (Unmarked SSS inputs). There exist positive constants , , and , together with a full event and a deleted-matrix event , with the following properties. The latter event is signed-permutation invariant, contains only full-rank matrices, and implies . Uniformly for ,
On ,
For every fixed and every such ,
Finally, for the iid square matrix,
uniformly on any sufficiently small subwindow of the displayed range.
Source and conventions. Apply [10] to . Their Lemmas 4.2 and 8.2 give (5.1). The deterministic comparison (5.2) follows from equation (27) in the proof of their Lemma 4.3. Their Lemma 8.1, equations (49)–(50), gives (5.3), and their Corollary 1.3 gives (5.4). The replacement estimate holds for each fixed deleted matrix in their regularity event. Decrease and and increase if needed, choosing the displayed closed threshold interval strictly inside the source’s open intervals. Take smaller also than the exponential constants in (2.5) and (2.7).
The event is their from Section 8.1. The singular-vector convention preceding (4.1) makes its spectral and coordinate-flatness conditions invariant. Its LCD condition is invariant as well, since signed permutations preserve the integer lattice and Euclidean distances. The event includes the fresh-vector projection conditions from their Section 4. These conditions depend on , so must be kept distinct from the deleted-matrix event .
Theorem 5.2 (Transfer for orbit-bounded marks). Let be measurable, set to zero on rank deficiency, and suppose
for every full-rank , where is fixed. Then there exist , , and a deterministic sequence , depending only on the entry law and , such that
The estimate is uniform over all marks satisfying (5.5). No concentration of around one is required.
Proof. We may suppose . Write from Proposition 5.1, choose
where is fixed, and put . Choose and . These choices leave fixed slack for all threshold factors below. All assertions are for sufficiently large .
Removing large and very small marks. Lemma 3.1, applied to and (2.5), yields
By monotonicity and Lemma 2.2,
For the coefficient, invariance of , Lemma 3.1, and Proposition 4.1 give
Indeed the large-mark part is bounded by , and the small-mark part by .
Unmarked regularity at the enlarged threshold. On , . Consequently,
Only the unmarked estimate (36) is used. Since and , its range is respected.
Let
The inequality gives . Conversely, (37), (43)–(46) give, for a fixed constant ,
Conditional replacement and linearization. For , the effective threshold lies between and . The fixed endpoint slack also permits every threshold with , so (38) applies pointwise throughout this interval. Its averaged error is bounded by . Lemma 4.3, used at thresholds , and show that
The threshold displacement costs because
Combining (45), (47), and (48) proves (41), for example with
The terms and are . None of the errors requires .
Remark 5.3 (Why the proof preserves exponential precision). The truncation uses an orbitwise bound inside , rather than the unconditional probability of . Thus its cost is . The bounded part loses . Taking slowly makes both errors even for . There is no additional requirement on the rank or condition number of the preconditioner.
Proposition 5.4 (Normalization of the intensity). The intensity in (30) satisfies .
Proof. Choose a fixed sufficiently large that and its bounded multiplicative perturbations lie inside both the transfer window and the scalar subwindow in (39). Apply Theorem 5.2 to the mark equal to one on full rank to get
It remains to compare with . For , the rank-one norm in (13) is at least precisely on , and . If and , then (13) implies , away from square singularity. Hence
Set and use (39). Both bounds equal . This proves the normalization without using the anisotropic theorem.
Proof of the main theorem
We first prove the inverse form, which also yields the rectangular-input applications.
Theorem 6.1 (Universal inverse amplification). Under (1.1), for every deterministic with , the failure probability
satisfies
in a window (1.4), uniformly in and . The inverse norm is assigned on square singularity.
Proof. The mark satisfies (5.5) with . By Theorem 5.2,
uniformly in in the transfer window. The normal line satisfies the coordinate input of Theorem 3.3 by Lemma 3.4. The invariant intensity has uniformly bounded norm by Proposition 4.1. Therefore Theorem 3.3 and Proposition 5.4 give
uniformly over , . Its boundedness also follows directly from orbit averaging. Apply (2.6) at thresholds . Choose the main window strictly inside the transfer window. The threshold displacements cost , while (2.7) costs . Setting
balances these errors. It is less than uniformly for large once . Gather the uniform vanishing coefficient and transfer errors into . Square and deleted singularity probabilities are absorbed into the exponential remainder, decreasing and then if necessary so that . This proves (6.1). □
Proof of Theorem 1.1. For nonsingular , with as in (1.2),
On singularity both events count failure. Theorem 6.1 proves (1.5). Uniform negligibility of its errors and the coefficient lower bound proved next give (1.6). □
The Gaussian coefficient and anisotropy profiles
Proposition 7.1 (Bounds and diffuse bias). Let with . For the functional defined in (1.3),
and
For a sequence , if and only if .
Proof. The upper bound is Jensen’s inequality. If are the eigenvalues of , concavity of the square root gives pointwise
Taking expectations gives the middle lower bound. If is the Gaussian polar decomposition, and are independent and . Thus .
Set . The sphere moment formulas give
The exact identity , together with , proves (7.2).
For necessity, implies in , hence in probability. By convexity,
Thus is uniformly integrable and . The variance formula implies . Sufficiency follows from (7.2).
Proof of Corollary 1.3. Combine (7.2) with Theorem 1.1.
Remark 7.2 (Effective ranks). For and ,
Thus divergence of and divergence of the operator stable rank are equivalent qualitative conditions. They encode different quantitative information. The coefficient bias in (7.2) is quadratic in , whereas the inverse residual is governed by .
Proposition 7.3 (General spectral profiles). Let the eigenvalues of be , with sum one. Suppose for every fixed , and put . Then
The series is almost surely finite. In particular, for equal surviving weights carrying all the mass, the limit is
Proof. Gaussian polar decomposition gives the exact identity
The prefactor tends to one. For fixed , the first weighted terms converge in distribution. The centered remainder has variance
Its mean tends to . Taking first and then identifies the limit of the quadratic sum as , because . The series has finite expectation. The square roots have second moment one and hence are uniformly integrable, proving (7.3). Formula (7.4) follows by integrating the chi density.
Remark 7.4 (Finite-dimensional normalization). The coefficient describes the dimension-asymptotic lower tail in Theorem 1.1. Exact fixed-dimensional coefficients can contain additional factors tending to one, as illustrated by the tall Gaussian model in Proposition 10.4.
Applications
Deterministic inputs and ellipsoidal amplification
Corollary 8.1. For any deterministic nonzero , uniformly in (1.4),
For one nonzero deterministic right-hand side ,
Proof. Normalize and apply Theorem 6.1. For a single column, use Proposition 7.3 with one weight equal to one.
The norm in (8.1) is the largest amplification of an input in the ellipsoid . Taking gives worst-case amplification, whereas choosing with orthonormal columns gives amplification restricted to a prescribed input subspace. The theorem also covers rank-deficient input maps.
Square sample covariance with arbitrary population anisotropy
Corollary 8.2. Let be deterministic positive definite and set
Then uniformly in (1.4),
Proof. Use and apply Theorem 1.1.
The conclusion holds for arbitrary positive-definite , with symmetric iid latent entries and thresholds in the stated shrinking window.
Independent random environments
Corollary 8.3 (Conditional universality). Let be a sigma-field independent of , and let be an almost surely invertible -measurable matrix. Put . For deterministic in (1.4), almost surely,
where and all error constants are deterministic and uniform in the environment.
Proof. Condition on and use the uniform bound of Theorem 1.1. The residual is bounded by and can be included in .
For two independent fresh iid matrices, also independent of the same environment, let be the corresponding failures with the same and threshold. Conditional independence and (59) imply
Indeed the covariance equals the variance of the conditional failure probability, and its ratio to differs uniformly from by . Thus concentration of the coefficient is sufficient for asymptotic decorrelation in this conditional-independence setting.
Independent exponential families
Corollary 8.4 (Independent extremes). Let with . Let , , be independent copies of , and let be deterministic invertible matrices. Define
If , then for each ,
and converges in distribution to an exponential random variable of rate .
Proof. The threshold lies in (4) for large . Theorem 1.1, uniformly in , gives probabilities with and . The accumulated exponential floor tends to zero because . Independence and the product of the Bernoulli probability-generating functions prove (8.6). The zero-count probability gives the minimum law.
Scope of the transfer principle
Theorems 3.3 and 5.2 have distinct roles. The former permits any invariant weight and requires no quantitative convergence rate relative to the rare-event threshold. The latter carries a measurable mark through an exponentially accurate geometric comparison using an orbitwise second-moment bound. Their combination applies whenever a rare-event problem produces these two structures.
The symmetry assumption is used for signed-permutation invariance. Removing it would require a replacement for the orbit identities, not merely a change in the scalar small-ball theorem. Dependence between a preconditioner and the deleted matrix also needs separate analysis. For example, with and , the random positive-definite matrix
has trace one and diffuse squared spectral mass, but . Thus diffuseness alone does not control an adapted mark.
For overlapping minors, an application would require both the conditional law of the fresh matrix and estimates for joint failures. General orthogonal rotations of non-Gaussian iid entries do not preserve independence, so these points must be checked separately. In particular, extending Corollary 8.4 to dependent families requires joint estimates beyond the independent-environment results proved here.
Conjectures and further directions
Theorem 1.1 expresses the leading lower-tail coefficient as a Haar functional of the normalized inverse ellipsoid. This geometric description suggests extensions both to larger hard-edge thresholds and to rectangular matrices. At larger thresholds, the natural comparison object is the full Gaussian least-singular-value distribution. For tall rectangular matrices, volumes of higher-dimensional orthogonal projections of the inverse ellipsoid suggest analogues of the square mean-width coefficient. The conjectures below formulate these extensions when the row and column dimensions differ by a fixed amount.
Correlated and spiked Gaussian hard-edge laws [3, 13, 7] and quantitative universality results for unpreconditioned iid matrices [11, 12] provide natural antecedents. Here the proposed comparison with the Gaussian model is uniform over all deterministic invertible preconditioners. For the rectangular coefficient asymptotic, this uniformity also concerns the Gaussian small-threshold remainder when the inverse profile approaches rank deficiency.
Universality from rare tails to the ordinary hard edge
Fix an integer . Let be an matrix with iid entries of law , and let have iid standard real Gaussian entries. Here denotes the least of the singular values, and equals zero on rank deficiency. For , retain the normalization from (1.2) and write
and define analogously.
Conjecture 10.1 (Uniform Gaussian comparison across hard-edge scales). Let be a fixed real, centered, variance-one, subgaussian random variable. Symmetry is not assumed. For each fixed , there exists such that, for every fixed ,
For and , the comparison in Conjecture 10.1 already follows from [10 Theorem 1.2], including for nonsymmetric entry laws, after choosing the exponential endpoint sufficiently small. For arbitrary , Theorem 1.1 establishes the symmetric square case only in the window (1.4). The conjecture extends this anisotropic comparison to ordinary hard-edge thresholds, nonsymmetric entries, and every fixed excess of columns. An additive comparison at ordinary thresholds would still leave an intermediate shrinking-threshold range unless its error bound overlaps with that of the present theorem.
The Gaussian reference must retain the full inverse profile. In the square case , its limiting distribution is [1, 11]. For a rank-one input, the reciprocal Gaussian amplification is exactly . More generally, the finite-rank identity in Proposition 10.3 below gives explicit benchmarks, while spiked Gaussian hard-edge laws provide further antecedents [7]. The role of in Theorem 1.1 is to describe the first-order rare-tail behavior.
The exponential endpoint must depend on the entry law and on . For example, for , coincidence of two rows of has probability and forces rank deficiency. Such atoms must be negligible compared with the Gaussian lower tail. In contrast with the tall model below, arbitrary inverse anisotropy does not make the wide Gaussian coefficient vanish.
For this wide model, the candidate dimension-asymptotic small-threshold coefficient is
The proposed asymptotic is a lower tail of order with coefficient as . Exact fixed-dimensional Gaussian coefficients may include further normalization factors. At it is the coefficient of this paper. At , exactly. For , the functional is convex and symmetric in the eigenvalues of , so concentrating inverse singular mass increases it in the majorization order, reversing the concave ordering at . A rank-one profile gives in the large- limit, matching the lower-tail coefficient of .
A rectangular intrinsic-volume hierarchy
Now let have size and iid entries of law , where is fixed and . The multiplier still acts on the rows. For trace-one , define
where the columns of form a Haar-uniform orthonormal -frame in . Thus .
Conjecture 10.2 (Intrinsic-volume coefficients at a rectangular hard edge). Assume (1.1). For each fixed , there are constants , with , and a deterministic , depending only on and the entry law, such that, uniformly over and
where and is zero on rank deficiency.
The coefficients have an intrinsic-volume interpretation. Let be the volume of the unit ball in , and let denote the intrinsic volume with the Steiner normalization. Kubota’s formula gives
See [2] for this normalization and ellipsoid formulas. The square coefficient is therefore the first member, corresponding to mean width. The conjecture would identify all fixed-order intrinsic volumes of the inverse ellipsoid as coefficients of real rectangular hard-edge tails.
The distinction between the two rectangular models is essential. One has , but for this coefficient can tend to zero. Indeed, Cauchy–Binet and Haar symmetry give
where is the th elementary symmetric polynomial. Jensen’s inequality then gives the upper bound on . If all inverse mass concentrates on limiting directions, the projected determinant vanishes. Under a finite-spike profile with weights and diffuse residual mass , the corresponding geometric limit is
This follows by the Gaussian limit of a fixed number of Haar rows and concentration of the diffuse remainder. The preceding determinant identity supplies uniform integrability of the square roots.
Consequently (10.4) gives a relative asymptotic when is bounded below. A relative comparison uniform over all tall preconditioners, without an additive floor, would be false for discrete entries. For example, duplicate columns give a positive rank-deficiency probability for Rademacher matrices. With and , the Gaussian lower block has full column rank almost surely, while as . The Gaussian failure probability therefore tends to zero, whereas the discrete singularity atom is unchanged.
Exact Gaussian identities
The following propositions give exact Gaussian identities that fix the normalizations in the two rectangular models.
Proposition 10.3 (Gaussian benchmarks). Let be as above. If has orthonormal columns and is a positive diagonal matrix, then
Here denotes the Moore–Penrose inverse. In particular, the reciprocal for , has law . Moreover, for , every , every invertible , and every ,
Proof. Put . The squared norm in (10.7) equals
Rotate to the first coordinate vectors and condition on the remaining Gaussian rows. Their orthogonal complement has dimension , and the projected first rows form an iid Gaussian matrix. The Schur complement identifies the leading block of with the inverse Gram matrix of this projection, proving (10.7).
For (10.8), has density proportional to on the positive-definite cone. The event in the survival probability is . Translation by multiplies its integral by . This is also a special case of the correlated Gaussian Wishart formulas in [13].
Proposition 10.4 (The tall Gaussian coefficient). Fix and . Let be an matrix with iid standard real Gaussian entries and uniform on . With and ,
Proof. Replacing by , we may assume . Put . The joint density of the ordered squared singular values of is proportional to
The substitution and , , shows that is exponential with rate , independently of . Choose the left singular frame with independent uniform column signs. It is Haar and independent of these variables. Set
For , take . The event is equivalent to , and hence to . Thus
The largest eigenvalue on the right is positive, since its Rayleigh quotient at the first coordinate vector is . The zero eigenvalue of is simple almost surely, so
If is a unit normal to the columns of , the Schur complement yields
In particular, , and the conditional probability divided by is bounded by . Dominated convergence therefore gives the coefficient . Conditional on , the first column of is uniform on the unit sphere of . Averaging its quadratic form gives the first equality in (70).
For the second equality, the ellipsoid formula in [2 Section 2] gives
Now apply (66) with , using and .
This identity determines the coefficient at fixed dimension and fixed preconditioner. A remainder bound uniform over preconditioners, as required by Conjecture 10.2, remains to be established.
Further inputs required
Conjecture 10.1 would require control of the bottom singular values jointly with their projections onto deterministic directions, and truncation of the remaining inverse spectrum. At ordinary thresholds several low singular modes can contribute. Removing symmetry also requires a substitute for signed-permutation averaging. For , the rarity scale is , so a direct extension of the present truncation needs higher-order mark control beyond the orbitwise second-moment identity.
Conjecture 10.2 additionally calls for comparison of a random -dimensional normal space and its volume mark at matrix-dependent thresholds. Together, the conjectures concern arbitrary deterministic population anisotropy throughout the real fixed-codimension least-singular-value problem. They make no assertion about the full eigenvalue process, proportional-aspect soft edges, or growing rectangular excess. Quantitative rates for the existing theorem remain a separate question: (27) pairs an Wasserstein bound for with the uniform bound for .
The two-coordinate magnitude limit
We give the fixed-coordinate consequence needed in Lemma 3.4, using the projected-column method of Nguyen–Vu [6 Section 6]. The sampling concentration estimate in their Theorem 6.1 is stated without dependence on the sample size. Since convergence in probability of the normalization is sufficient here, we replace that step by the finite-population variance bound below. The other inputs are their maximal-coordinate delocalization, their frame Berry–Esseen estimate, and polynomial flatness of the omitted-column complement. All references in this appendix to results of [6] use the published version.
Lemma A.1 (Sampled normalization). Let , , and with and . Let be a uniformly chosen -element subset of containing 1 and 2, and set
Then
In particular, in probability whenever .
Proof. Put . The remaining entries are sampled without replacement from a population of size , and
This gives the first bound. The finite-population variance formula, followed by , gives
The conclusion follows from these bounds and Chebyshev’s inequality.
Proof of Lemma 3.4. Write , so is an iid matrix, and let be its unit normal on full rank. The rectangular estimate [9 Theorem 1.1] makes rank deficiency exponentially unlikely. Coordinate delocalization [6 Theorem 1.4, equation (6)] supplies a constant such that
Sampling and normalization. Independently of , choose a uniform -element set containing 1 and 2, ordered with these indices first. The parameter will be a sufficiently small positive power of . On the event in (A.2) with the reverse inequality, apply Lemma A.1 conditionally to and . Since , it follows that
in probability. (A.3)
This estimate will be used after the comparison below. We do not condition the retained entries on the event in (A.2) or on a restriction on .
Projection and Gaussian comparison. Let be the span of the columns of indexed by . With exponentially high probability, those columns have full column rank, so . The flatness estimate [6 Lemma 6.3] gives a constant such that, for sufficiently small polynomial choices of , with probability over and the omitted columns,
Choose a fixed with small enough to meet that estimate, and put
The constant in the flatness estimate can be taken independently of this sufficiently small choice of .
Condition only on and an omitted-column realization satisfying full column rank and (A.4). Choose an orthonormal basis of , and let be the corresponding coordinate projection. Thus . Set , an matrix. The entries of are still independent copies of the original entry variable under this conditioning.
For and , let be the vectorization of the matrix . Then
The Gaussian replacement of this normalized frame is therefore the vectorization of an iid standard Gaussian matrix of the same dimensions. For any measurable set in the matrix space, the frame Berry–Esseen estimate [6 Lemma 2.5] gives
where is the -neighborhood of the boundary in the entrywise maximum norm, and
The constants are uniform over the conditioned realizations.
Control of the event boundary. For fixed , let consist of full-row-rank matrices whose unit kernel vector , defined up to sign, satisfies for . The Gaussian rectangular bound gives
where denotes its smallest nonzero singular value on full row rank. Suppose and differs from by at most in each entry. Then , so has full row rank. If is its unit kernel vector, projection onto the row space of gives
After choosing the sign of a unit vector , we obtain
The Gaussian kernel line is Haar. Each scaled coordinate magnitude of a Haar vector in has a density bounded uniformly for sufficiently large . Consequently,
Here . Combining this estimate with (A.5), and then adding the probability of omitted-column irregularity, proves
One may define arbitrarily on failure of full row rank. When and the omitted block both have full rank, the rank decomposition under projection implies . Their exceptional probabilities are exponentially small.
Return to the original coordinates. On this full-rank event, : otherwise the full-column-rank omitted block would annihilate the nonzero vector . Projecting shows that the kernel line of is spanned by , so up to sign. Therefore
Equations (A.3) and (A.7), together with Slutsky’s theorem, give the required limit. Independence between and is unnecessary. Finally, the Haar-line extension on rank deficiency changes none of these limits.
References
References
- [1]A. Edelman, Eigenvalues and condition numbers of random matrices, SIAM J. Matrix Anal. Appl. 9 (1988), no. 4, 543–560. doi:10.1137/0609045. MR0964668.DOI
- [2]A. Gusakova, E. Spodarev, and D. Zaporozhets, Intrinsic volumes of ellipsoids, Zap. Nauchn. Sem. POMI 515 (2022), 121–140. arXiv:2206.14002. MR4527616.arxiv.org/abs/2206.14002
- [3]W. Hachem, A. Hardy, and J. Najim, Large complex correlated Wishart matrices: Fluctuations and asymptotic independence at the edges, Ann. Probab. 44 (2016), no. 3, 2264–2348. doi:10.1214/15-AOP1022. MR3502605.DOI
- [4]C. E. Häggblom, Gaussian critical-threshold instability in real phase retrieval, preprint (2026), arXiv:2609.28198. The lemma numbering cited here refers to arXiv:2609.28198v1.arxiv.org/abs/2609.28198
- [5]B. Letwin and A. R. Polavarapu, On the smallest singular value of the product of random and deterministic matrices, preprint (2026), arXiv:2607.06785.arxiv.org/abs/2607.06785
- [6]H. H. Nguyen and V. H. Vu, Normal vector of a random hyperplane, Int. Math. Res. Not. IMRN 2018 (2018), no. 6, 1754–1778. doi:10.1093/imrn/rnw273. Preprint: arXiv:1604.04897. References to Lemmas 2.5 and 6.3 and to the results in Section 6 refer to the published version.
- [7]J. A. Ramírez and B. Rider, Spiking the random matrix hard edge, Probab. Theory Related Fields 169 (2017), 425–467. doi:10.1007/s00440-016-0733-1. MR3704773.
- [8]M. Rudelson and R. Vershynin, The Littlewood–Offord problem and invertibility of random matrices, Adv. Math. 218 (2008), no. 2, 600–633. doi:10.1016/j.aim.2008.01.010. MR2407948.DOI
- [9]M. Rudelson and R. Vershynin, Smallest singular value of a random rectangular matrix, Comm. Pure Appl. Math. 62 (2009), no. 12, 1707–1739. doi:10.1002/cpa.20294. MR2569075.DOI
- [10]A. Sah, J. Sahasrabudhe, and M. Sawhney, On the Spielman–Teng conjecture, Geom. Funct. Anal. 35 (2025), 633–671. doi:10.1007/s00039-025-00707-z. The lemma numbering used here agrees with arXiv:2405.20308v2.DOI
- [11]T. Tao and V. Vu, Random matrices: The distribution of the smallest singular values, Geom. Funct. Anal. 20 (2010), no. 1, 260–297. doi:10.1007/s00039-010-0057-8. MR2647142.DOI
- [12]H. Wang, Optimal smoothed analysis and quantitative universality for the smallest singular value of random matrices, preprint (2022), arXiv:2211.03975.arxiv.org/abs/2211.03975
- [13]T. Wirtz and T. Guhr, Distribution of the smallest eigenvalue in complex and real correlated Wishart ensembles, J. Phys. A: Math. Theor. 47 (2014), no. 7, 075004. doi:10.1088/1751-8113/47/7/075004.DOI