betaFeedback

WorksMathematics

Ultra-hard-edge universality under deterministic preconditioning

Abstract

Let AnA_n be an n×nn\times n real matrix with iid entries distributed as a fixed centered, variance-one, symmetric subgaussian random variable. For deterministic invertible MnM_n, set hn=∥Mn−1∥HSh_n=\|M_n^{-1}\|_{\mathrm{HS}} and Bn=Mn−1/hnB_n=M_n^{-1}/h_n. We prove P(smin⁡(MnAn)≤ε/hn)=(1+o(1))εΘn(Bn)\mathbb{P}(s_{\min}(M_nA_n)\le\varepsilon/h_n) =(1+o(1))\varepsilonΘ_n(B_n) as n→∞n\to\infty, uniformly over MnM_n and e−cn≤ε≤n−Ce^{-cn}\le\varepsilon\le n^{-C}. Here Θn(B)=EnU⊤BB⊤UΘ_n(B)=\mathbb{E}\sqrt{nU^\top BB^\top U}, where UU is uniform on Sn−1S^{n-1}. The constants c>0c>0 and C>1C>1 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, ξ\xi is a fixed real random variable satisfying

Eξ=0,Eξ2=1,ξ=d−ξ,∥ξ∥ψ2<∞.(1)\mathbb{E}\xi=0,\qquad\mathbb{E}\xi^2=1,\qquad\xi\mathrel{\overset{d}{=}}-\xi,\qquad\|\xi\|_{\psi_2}<\infty. \tag*{(1)}

Let AnA_n be an n×nn\times n matrix whose entries are independent copies of ξ\xi. For a deterministic invertible matrix MM, set

h(M)=∥M−1∥HS,B=M−1h(M),H=BBT,Tr⁡H=1.(2)h(M)=\|M^{-1}\|_{\mathrm{HS}},\qquad B=\frac{M^{-1}}{h(M)},\qquad H=BB^{\mathsf T},\qquad\operatorname{Tr}H=1. \tag*{(2)}

Here ∥T∥HS2=Tr⁡(TTT)\|T\|_{\mathrm{HS}}^2=\operatorname{Tr}(TT^{\mathsf T}) is the squared Hilbert–Schmidt norm. For H⪰0H\succeq0 with Tr⁡H=1\operatorname{Tr}H=1 and B∈Rn×mB\in\mathbb{R}^{n\times m} with ∥B∥HS=1\|B\|_{\mathrm{HS}}=1, define

ϑn(H):=EU∼Haar⁡(Sn−1)nUTHU,Θn(B):=ϑn(BBT).(3)\vartheta_n(H):=\mathbb{E}_{U\sim\operatorname{Haar}(\mathbb{S}^{n-1})}\sqrt{nU^{\mathsf T}HU},\qquad\Theta_n(B):=\vartheta_n(BB^{\mathsf T}). \tag*{(3)}

Thus Θn(B)\Theta_n(B) depends only on the eigenvalues of BBTBB^{\mathsf T}. For any nonzero deterministic K∈Rn×mK\in\mathbb{R}^{n\times m},

Θn(K∥K∥HS)=n∥K∥HSEU∥KTU∥2.\Theta_n\left(\frac{K}{\|K\|_{\mathrm{HS}}}\right)=\frac{\sqrt{n}}{\|K\|_{\mathrm{HS}}}\mathbb{E}_U\|K^{\mathsf T}U\|_2.

For M=IM=I, one has h(I)=nh(I)=\sqrt{n}, H=I/nH=I/n, and Θn(I/n)=1\Theta_n(I/\sqrt{n})=1. The ordinary hard-edge scale is then n−1/2n^{-1/2}. We use the term ultra-hard edge for the smaller thresholds ε/h(M)\varepsilon/h(M) considered below, where ε\varepsilon 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 0<c0<c10<c_0<c_1, C0>1C_0>1, C<∞C<\infty, and a deterministic sequence rn→0r_n\to0, depending only on the entry law, such that for every deterministic invertible n×nn\times n matrix MM and every

e−c0n≤ε≤n−C0,(4)e^{-c_0n}\leq\varepsilon\leq n^{-C_0}, \tag*{(4)}

with BB defined in (2),

∣P(smin⁡(MAn)≤ε∥M−1∥HS)−εΘn(B)∣≤Cε[rn+(εn∥B∥op2)1/3]+Ce−c1n.(5)\left|\mathbb{P}\left(s_{\min}(MA_n)\leq\frac{\varepsilon}{\|M^{-1}\|_{\mathrm{HS}}}\right)-\varepsilon\Theta_n(B)\right|\leq C\varepsilon\left[r_n+(\varepsilon n\|B\|_{\mathrm{op}}^2)^{1/3}\right]+Ce^{-c_1n}. \tag*{(5)}

All bounds are uniform over MM. In particular, throughout (4),

P(smin⁡(MAn)≤ε∥M−1∥HS)=(1+o(1))εΘn(B).(6)\mathbb{P}\left(s_{\min}(MA_n)\leq\frac{\varepsilon}{\|M^{-1}\|_{\mathrm{HS}}}\right)=(1+o(1))\varepsilon\Theta_n(B). \tag*{(6)}

The coefficient is uniformly nondegenerate:

2π≤Θn(B)≤1.(7)\sqrt{\frac{2}{\pi}}\leq\Theta_n(B)\leq1. \tag*{(7)}

We prove this and more precise profile statements in § 7. Since ∥B∥op≤1\|B\|_{\mathrm{op}}\leq1, the explicit residual in (5) is uniformly negligible in (4). The inequality c0<c1c_0<c_1 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 GnG_n have iid standard real Gaussian entries. Under the assumptions and in a common window of the form (4),

P(smin⁡(MAn)≤ε/∥M−1∥HS)P(smin⁡(MGn)≤ε/∥M−1∥HS)⟶1(8)\frac{\mathbb{P}\left(s_{\min}(MA_n)\leq\varepsilon/\|M^{-1}\|_{\mathrm{HS}}\right)} {\mathbb{P}\left(s_{\min}(MG_n)\leq\varepsilon/\|M^{-1}\|_{\mathrm{HS}}\right)} \longrightarrow1 \tag*{(8)}

uniformly over deterministic invertible MM and admissible ε\varepsilon.

Proof. Apply Theorem 1.1 to both entry laws, shrink the window to their intersection, and use (7).

Diffuse and concentrated inverse profiles

Define

Δ4(M)2=Tr⁡((BB⊤)2).(9)\Delta_4(M)^2 = \operatorname{Tr}\left((BB^\top)^2\right). \tag*{(9)}

The quantity Δ4(M)\Delta_4(M) 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 Δ4(Mn)→0\Delta_4(M_n) \to0, then uniformly in (1.4),

P(smin⁡(MnAn)≤ε∥Mn−1∥HS)=(1+o(1))ε.\mathbb{P}\left(s_{\min}(M_n A_n) \le\frac{\varepsilon}{\lVert M_n^{-1}\rVert_{\mathrm{HS}}}\right) = (1+o(1))\varepsilon.

More precisely, (1.5) remains valid with εΘn(B)\varepsilon\Theta_n(B) replaced by ε\varepsilon, provided εΔ4(Mn)2\varepsilon\Delta_4(M_n)^2 is added to its right-hand side.

At the other extreme, when the normalized inverse singular mass approaches rank one, the coefficient tends to 2/π\sqrt{2/\pi}. More generally, if finitely many weights λ1,…,λr\lambda_1,\ldots,\lambda_r survive and the residual mass d=1−∑j=1rλjd=1-\sum_{j=1}^{r}\lambda_j is diffuse, the coefficient tends to

Ed+∑j=1rλjgj2,gj independent N(0,1).(10)\mathbb{E}\sqrt{d+\sum_{j=1}^{r}\lambda_j g_j^2}, \qquad g_j\text{ independent }N(0,1). \tag*{(10)}

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 ∥M−1∥HS−1\lVert M^{-1}\rVert_{\mathrm{HS}}^{-1} 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 EU∥K⊤U∥2\mathbb{E}_U\lVert K^\top U\rVert_2, where UU is Haar-uniform on Sn−1S^{n-1}. For a nonzero $n\times m input matrix KK, this average satisfies

EU∥KTU∥2=∥K∥HSnΘn(K∥K∥HS).\mathbb{E}_U\left\|K^{\mathsf{T}}U\right\|_2=\frac{\|K\|_{\mathrm{HS}}}{\sqrt{n}}\Theta_n\left(\frac{K}{\|K\|_{\mathrm{HS}}}\right).

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

aB(C)=n∥BTv(C)∥2,a_B(C)=\sqrt{n}\left\|B^{\mathsf{T}}v(C)\right\|_2,

where CC is obtained by deleting one column of AnA_n and v(C)v(C) 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 aBa_B, averaging aB(PC)2a_B(PC)^2 over signed permutation matrices PP gives exactly one. This identity permits truncation at aB≤La_B \le L with rare-event cost O(ε/L)O(\varepsilon/L). On the retained set, the unmarked estimates apply at the enlarged threshold εL\varepsilon L. The resulting coefficient is E[aB(C)Λn(C)]\mathbb{E}[a_B(C)\Lambda_n(C)], where Λn\Lambda_n is a signed-permutation-invariant spectral intensity.

The second ingredient is the weighted coefficient comparison, Theorem 3.3. It identifies this expectation uniformly in BB, using an LpL^p bound for Λn\Lambda_n with 1<p<21 < p < 2 and weak convergence of two coordinate magnitudes of v(C)v(C). 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 Θn(B)\Theta_n(B).

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 ∥T∥op\|T\|_{\mathrm{op}} for the Euclidean operator norm and use the subgaussian norm

∥ξ∥ψ2:=inf⁡{s>0:Eexp⁡(ξ2/s2)≤2}.\|\xi\|_{\psi_2}:=\inf\{s>0:\mathbb{E}\exp(\xi^2/s^2)\le2\}.

Constants may depend on the fixed entry law. The notation o(1)o(1) always denotes a deterministic quantity tending to zero. When matrices or thresholds vary, uniformity is stated explicitly. Singular values of a deleted n×(n−1)n \times(n-1) matrix are ordered increasingly as s1(C)≤⋯≤sn−1(C)s_1(C) \le\cdots\le s_{n-1}(C). Square least singular values are denoted by smin⁡s_{\min}. 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 O(e−cn)O(e^{-cn}). In inverse-amplification statements, we count a singular AnA_n as a failure and assign its inverse norm the value +∞+\infty.

The logarithmic cutoffs and regularity events below are defined for n≥n0n \ge n_0, with n0n_0 sufficiently large. Uniform moment bounds for the associated intensities are understood over this range. The value of n0n_0 may be increased in the asymptotic arguments.

The inverse geometry

Write A=[C X]A=[C\ X], where C∈Rn×(n−1)C \in\mathbb{R}^{n \times(n-1)} and X∈RnX \in\mathbb{R}^n. On {rank⁡C=n−1}\{\operatorname{rank} C=n-1\} choose QQ with orthonormal columns spanning col⁡C\operatorname{col} C, and choose a measurable unit vector v∈ker⁡CTv \in\ker C^\mathsf{T}. Put

R=QTC,g=QTX,x=vTX.R=Q^\mathsf{T}C,\qquad g=Q^\mathsf{T}X,\qquad x=v^\mathsf{T}X.

Then (Q,v)TA=(Rg0x)(Q,v)^\mathsf{T}A=\begin{pmatrix}R & g\\ 0 & x\end{pmatrix}.

Proposition 2.1 (One-column decomposition). For any deterministic B∈Rn×mB \in\mathbb{R}^{n \times m}, on {rank⁡C=n−1,x≠0}\{\operatorname{rank} C=n-1,x\ne0\},

A−1B=EC+x−1wC(X)vTB,EC=(R−1QTB0),wC(X)=(−R−1QTX1).(11)A^{-1}B=E_C+x^{-1}w_C(X)v^\mathsf{T}B,\qquad E_C=\begin{pmatrix}R^{-1}Q^\mathsf{T}B\\0\end{pmatrix},\qquad w_C(X)=\begin{pmatrix}-R^{-1}Q^\mathsf{T}X\\1\end{pmatrix}. \tag*{(11)}

If

ΞC(X)=1+∥R−1QTX∥22,DC=∥EC∥op⁡,(12)\Xi_C(X)=\sqrt{1+\lVert R^{-1}Q^\mathsf{T}X\rVert_2^2},\qquad D_C=\lVert E_C\rVert_{\operatorname{op}}, \tag*{(12)}

then

∣∥A−1B∥op⁡−ΞC(X)∥BTv∥2∣vTX∣∣≤DC≤∥B∥op⁡s1(C).(13)\left|\lVert A^{-1}B\rVert_{\operatorname{op}}-\frac{\Xi_C(X)\lVert B^\mathsf{T}v\rVert_2}{|v^\mathsf{T}X|}\right|\le D_C\le\frac{\lVert B\rVert_{\operatorname{op}}}{s_1(C)}. \tag*{(13)}

Proof. Block inversion proves (2.1). The second term has rank at most one, with norm

∣x∣−1∥wC(X)∥2∥BTv∥2.|x|^{-1}\lVert w_C(X)\rVert_2\lVert B^\mathsf{T}v\rVert_2.

Since C=QRC=QR, the singular values of RR are the nonzero singular values of CC. The triangle inequality gives the remaining assertions. Also R−1QT=C†R^{-1}Q^\mathsf{T}=C^\dagger, so ΞC(X)=1+∥C†X∥22\Xi_C(X)=\sqrt{1+\lVert C^\dagger X\rVert_2^2} is independent of the choice of QQ.

For subsequent unconditional probability estimates, set DC=+∞D_C=+\infty on {rank⁡C<n−1}\{\operatorname{rank} C<n-1\}.

For a nonnegative CC-measurable mark aa, let

Tta={rank⁡C=n−1, ∣vTX∣≤ta(C)n−1/2ΞC(X)},Ts=Ts⋅1.(14)T_{ta}=\{\operatorname{rank} C=n-1,\ |v^\mathsf{T}X|\le ta(C)n^{-1/2}\Xi_C(X)\},\qquad T_s=T_{s\cdot1}. \tag*{(14)}

Also put Ss={smin⁡(A)≤s/n}S_s=\{s_{\min}(A)\le s/\sqrt{n}\}.

Lemma 2.2 (Exact event inclusion). For every s≥0s\ge0, Ts⊂SsT_s\subset S_s. If AA is iid as in (1.1), then

P(Ts)≤Cs+Ce−cn,s≥0.(15)\mathbb{P}(T_s)\le Cs+Ce^{-cn},\qquad s\ge0. \tag*{(15)}

Proof. The vector wC(X)w_C(X) from (2.1) satisfies AwC(X)=(vTX)vAw_C(X)=(v^\mathsf{T}X)v and ∥wC(X)∥2=ΞC(X)\lVert w_C(X)\rVert_2=\Xi_C(X). Hence smin⁡(A)≤∣vTX∣/ΞC(X)s_{\min}(A)\le|v^\mathsf{T}X|/\Xi_C(X), including when vTX=0v^\mathsf{T}X=0. The probability bound follows from the square least-singular-value estimate [8].

For ∥B∥HS=1\lVert B\rVert_{\mathrm{HS}}=1 define aB(C)=n∥BTv(C)∥2a_B(C)=\sqrt{n}\lVert B^{\mathsf{T}}v(C)\rVert_2 on full rank, and zero otherwise. Proposition 2.1 gives, for 0<δ<10<\delta<1, outside singularity and {DC>δ/ε}\{D_C>\delta/\varepsilon\},

TεaB/(1+δ)⊂{∥A−1B∥op≥ε−1}⊂TεaB/(1−δ).(16)T_{\varepsilon a_B/(1+\delta)}\subset\{\lVert A^{-1}B\rVert_{\mathrm{op}}\geq\varepsilon^{-1}\}\subset T_{\varepsilon a_B/(1-\delta)}. \tag*{(16)}

The rectangular estimate [9] gives

P(DC>δ/ε)≤Cε2n∥B∥op2δ2+Ce−cn.(17)\mathbb{P}(D_C>\delta/\varepsilon)\leq C\frac{\varepsilon^2 n\lVert B\rVert_{\mathrm{op}}^2}{\delta^2}+Ce^{-cn}. \tag*{(17)}

Signed-permutation averaging

Let Gn\mathcal{G}_n denote the finite group of signed permutation matrices of order nn. All averages over P∈GnP\in\mathcal{G}_n use its uniform probability measure. For fixed z∈Rnz\in\mathbb{R}^n,

EPPzzTPT=∥z∥22nI.(18)\mathbb{E}_P Pzz^{\mathsf{T}}P^{\mathsf{T}}=\frac{\lVert z\rVert_2^2}{n}I. \tag*{(18)}

Under (1.1), (PC,PX)(PC,PX) has the same law as (C,X)(C,X), while the normal line transforms as v(PC)v(PC)T=Pv(C)v(C)TPTv(PC)v(PC)^{\mathsf{T}}=Pv(C)v(C)^{\mathsf{T}}P^{\mathsf{T}}.

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 ZZ with invariant law. Let (Fs)s≥0(F_s)_{s\geq0} be increasing invariant events satisfying

P(Fs)≤Ks+b,s≥0.\mathbb{P}(F_s)\leq Ks+b,\qquad s\geq0.

Suppose a(Z)≥0a(Z)\geq0 satisfies EPa(PZ)2≤Ka\mathbb{E}_P a(PZ)^2\leq K_a for every ZZ. Then, for t>0t>0 and L≥1L\geq1,

P(Fta(Z)∩{a(Z)>L})≤CKKatL+CKabL2.(19)\mathbb{P}(F_{ta(Z)}\cap\{a(Z)>L\})\leq CK K_a\frac{t}{L}+CK_a\frac{b}{L^2}. \tag*{(19)}

If W(Z)≥0W(Z)\geq0 is invariant and integrable, then

E[aW1{a>L}]≤KaLEW.(20)\mathbb{E}[aW1_{\{a>L\}}]\leq\frac{K_a}{L}\mathbb{E}W. \tag*{(20)}

Proof. Split {a>L}\{a>L\} into the intervals (2jL,2j+1L](2^jL,2^{j+1}L], j≥0j\geq0. Monotonicity, invariance of FsF_s, and Markov’s inequality on each orbit give

P(Fta∩{2jL<a≤2j+1L})≤E[1Ft2j+1LEP1{a(PZ)>2jL}]≤Ka22jL2P(Ft2j+1L).\mathbb{P}(F_{ta}\cap\{2^jL<a\leq2^{j+1}L\})\leq\mathbb{E}\left[1_{F_{t2^{j+1}L}}\mathbb{E}_P1_{\{a(PZ)>2^jL\}}\right] \leq\frac{K_a}{2^{2j}L^2}\mathbb{P}(F_{t2^{j+1}L}).

Summation gives (3.2). For the second assertion, average over the group and use a1{a>L}≤a2/La1_{\{a>L\}}\leq a^2/L.

For the norm mark, (3.1) gives the exact identity

EPaB(PC)2=1,rank⁡C=n−1,(21)\mathbb{E}_P a_B(PC)^2=1,\qquad\operatorname{rank}C=n-1, \tag*{(21)}

and consequently EaB≤1\mathbb{E}a_B\leq1. The full events TsT_s are invariant under simultaneous signed permutations of CC and XX. Lemmas 2.2 and 3.1 therefore imply

P(TtaB∩{aB>L})≤CtL+Ce−cn.(22)\mathbb{P}(T_{ta_B}\cap\{a_B>L\})\leq C\frac{t}{L}+Ce^{-cn}. \tag*{(22)}

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 v∈Sn−1v \in S^{n-1} and ∥B∥HS=1\lVert B\rVert_{\mathrm{HS}}=1, put

FB(v)=n∥BTv∥2,F‾B(v)=EPFB(Pv),μv=1n∑i=1nδn∣vi∣.F_B(v)=\sqrt{n}\lVert B^{\mathsf T}v\rVert_2,\qquad\overline{F}_B(v)=\mathbb{E}_P F_B(Pv),\qquad\mu_v=\frac{1}{n}\sum_{i=1}^{n}\delta_{\sqrt{n}|v_i|}.

Write W2W_2 for the quadratic Wasserstein distance on probability measures on [0,∞)[0,\infty).

Lemma 3.2 (Orbit transport). For all u,v∈Sn−1u,v\in S^{n-1} and all rectangular BB with ∥B∥HS=1\lVert B\rVert_{\mathrm{HS}}=1,

∣F‾B(v)−F‾B(u)∣≤W2(μv,μu).(23)\left|\overline{F}_B(v)-\overline{F}_B(u)\right|\le W_2(\mu_v,\mu_u). \tag*{(23)}

Moreover 0≤F‾B(v)≤10\le\overline{F}_B(v)\le1.

Proof. Let v↑,u↑v^\uparrow,u^\uparrow be the increasingly sorted coordinate magnitudes. They represent the same signed-permutation orbits as v,uv,u. Couple their orbit averages using the same PP, and put d=n(v↑−u↑)d=\sqrt{n}(v^\uparrow-u^\uparrow). The reverse triangle inequality, Cauchy–Schwarz, and (18) give

∣F‾B(v)−F‾B(u)∣≤EP∥BTPd∥2≤(∥d∥22nTr⁡BBT)1/2.\left|\overline{F}_B(v)-\overline{F}_B(u)\right|\le\mathbb{E}_P\left\lVert B^{\mathsf T}Pd\right\rVert_2\le\left(\frac{\lVert d\rVert_2^2}{n}\operatorname{Tr}BB^{\mathsf T}\right)^{1/2}.

Increasing matching is the optimal quadratic coupling of the two equally weighted one-dimensional empirical measures, so the last expression is W2(μv,μu)W_2(\mu_v,\mu_u). Finally EPFB(Pv)2=1\mathbb{E}_P F_B(Pv)^2=1, and Jensen’s inequality gives the upper bound.

Theorem 3.3 (Weighted coefficient comparison). Let (vn,Wn)(v_n,W_n) be a random unit line and a real random variable whose joint law is invariant under (vn,Wn)↦(Pvn,Wn)(v_n,W_n)\mapsto(Pv_n,W_n), P∈GnP\in G_n. Suppose

(n∣vn,1∣,n∣vn,2∣)⟹(∣g1∣,∣g2∣),(24)(\sqrt{n}|v_{n,1}|,\sqrt{n}|v_{n,2}|)\Longrightarrow(|g_1|,|g_2|), \tag*{(24)}

where g1,g2g_1,g_2 are independent standard Gaussians, and suppose sup⁡n∥Wn∥Lp<∞\sup_n\lVert W_n\rVert_{L^p}<\infty for some p>1p>1. Then

sup⁡m≥1 sup⁡B∈Rn×m∥B∥HS=1∣E[FB(vn)Wn]−Θn(B)EWn∣⟶0.(25)\sup_{m\ge1}\ \sup_{\substack{B\in\mathbb{R}^{n\times m}\\ \lVert B\rVert_{\mathrm{HS}}=1}}\left|\mathbb{E}[F_B(v_n)W_n]-\Theta_n(B)\mathbb{E}W_n\right|\longrightarrow0. \tag*{(25)}

No independence between vnv_n and WnW_n is assumed.

Proof. Let ν=L(∣g1∣)\nu=\mathcal{L}(|g_1|). For bounded continuous ff, exchangeability and (24) show that

E∫f dμvn→∫f dν,Var⁡(∫f dμvn)→0.\mathbb{E}\int f\,\mathrm{d}\mu_{v_n}\to\int f\,\mathrm{d}\nu,\qquad\operatorname{Var}\left(\int f\,\mathrm{d}\mu_{v_n}\right)\to0.

Applying this to a countable convergence-determining class gives μvn⇒ν\mu_{v_n}\Rightarrow\nu in probability. Both measures have second moment exactly one. The characterization of W2W_2 convergence by weak convergence and convergence of second moments therefore yields

W2(μvn,ν)→0in probability.(26)W_2(\mu_{v_n},\nu)\to0\quad\text{in probability}. \tag*{(26)}

This implication can also be applied along almost surely convergent subsubsequences. Since the measures are supported on [0,∞)[0,\infty) and have second moment one, their squared W2W_2 distance is at most two. Thus (26) holds in every fixed LqL^q, q<∞q<\infty.

For an independent Haar vector UnU_n, the same assertion follows by writing Un=g/∥g∥2U_n=g/\lVert g\rVert_2 and applying the law of large numbers. By Lemma 3.2,

sup⁡B∣F‾B(vn)−Θn(B)∣≤W2(μvn,ν)+EW2(μUn,ν).(27)\sup_B\left|\overline{F}_B(v_n)-\Theta_n(B)\right|\le W_2(\mu_{v_n},\nu)+\mathbb{E}W_2(\mu_{U_n},\nu). \tag*{(27)}

Here EFB‾(Un)=Θn(B)\mathbb{E}\overline{F_B}(U_n)=\Theta_n(B). Joint invariance gives

E[FB‾(vn)Wn]=E[FB‾(vn)Wn].\mathbb{E}[\overline{F_B}(v_n)W_n]=\mathbb{E}[\overline{F_B}(v_n)W_n].

Hölder’s inequality with q=p/(p−1)q=p/(p-1) and (27) finish the proof. All bounds are independent of BB and of its number of columns.

Lemma 3.4 (Two coordinate magnitudes). Let CnC_n be an n×(n−1)n\times(n-1) matrix with iid centered, variance-one, subgaussian entries. On full rank, let vnv_n be a unit vector spanning ker⁡CnT\ker C_n^{\mathsf T} and on rank deficiency extend its line by an independent Haar line. Then

(n∣vn,1∣,n∣vn,2∣)⟹(∣g1∣,∣g2∣),(\sqrt{n}\lvert v_{n,1}\rvert,\sqrt{n}\lvert v_{n,2}\rvert)\Longrightarrow(\lvert g_1\rvert,\lvert g_2\rvert),

where g1,g2g_1,g_2 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 CTCC^{\mathsf T}C as a function of that matrix alone, and form the corresponding left singular vectors uk=Cqk/sk(C)u_k=Cq_k/s_k(C). On full rank these obey uk(PC)=Puk(C)u_k(PC)=Pu_k(C), including with a fixed measurable choice inside repeated eigenspaces. This convention makes all coordinate-flatness and truncated spectral conditions below invariant under C↦PCC\mapsto PC.

Put ℓ=⌊log⁡n⌋\ell=\lfloor\sqrt{\log n}\rfloor and

χC(X)2=∑k=1ℓ⟨uk,X⟩2sk(C)2.(28)\chi_C(X)^2=\sum_{k=1}^{\ell}\frac{\langle u_k,X\rangle^2}{s_k(C)^2}. \tag*{(28)}

For G1,…,GℓG_1,\ldots,G_\ell iid N(0,1)N(0,1), let

YC=(1n∑k=1ℓGk2sk(C)2)1/2.(29)Y_C=\left(\frac{1}{n}\sum_{k=1}^{\ell}\frac{G_k^2}{s_k(C)^2}\right)^{1/2}. \tag*{(29)}

We shall use the SSS deleted-matrix event QnQ_n specified in Proposition 5.1. In particular it is invariant and implies s1(C)≥n−1/2(log⁡n)−3s_1(C)\ge n^{-1/2}(\log n)^{-3}. Define

Λn(C)=1Qn2πEGYC.(30)\Lambda_n(C)=\mathbf{1}_{Q_n}\sqrt{\frac{2}{\pi}}\mathbb{E}_G Y_C. \tag*{(30)}

This definition sets Λn=0\Lambda_n=0 off QnQ_n, including on rank deficiency. It is independent of the rare-event threshold and of the preconditioner.

Proposition 4.1 (Subcritical intensity moments). For every 0<p<20<p<2,

sup⁡n≥n0EΛn(C)p<∞.(31)\sup_{n\ge n_0}\mathbb{E}\Lambda_n(C)^p<\infty. \tag*{(31)}

The same conclusion holds if QnQ_n is replaced by any event contained in {s1(C)≥n−1/2(log⁡n)−A}\{s_1(C)\ge n^{-1/2}(\log n)^{-A}\}, for fixed A>0A>0.

Proof. It suffices to consider 1<p<21 < p < 2. Jensen’s inequality and subadditivity of the power p/2p/2 give

Λnp≤Cpn−p/2∑k=1ℓ1Qnsk(C)−p.(32)\Lambda_n^p \le C_p n^{-p/2} \sum_{k=1}^{\ell} \mathbf{1}_{Q_n}s_k(C)^{-p}. \tag*{(32)}

We claim

E[1Qnsk(C)−p]≤Cpnp/2k−p,1≤k≤ℓ.(33)\mathbb{E}\left[\mathbf{1}_{Q_n}s_k(C)^{-p}\right] \le C_p n^{p/2} k^{-p}, \qquad1 \le k \le\ell. \tag*{(33)}

Let CkC_k consist of the first n−kn-k columns of CC. Eigenvalue interlacing for the principal submatrix CkTCkC_k^{\mathsf{T}}C_k of CTCC^{\mathsf{T}}C gives s1(Ck)≤sk(C)s_1(C_k) \le s_k(C). By the rectangular estimate [9 Theorem 1.1],

P(s1(Ck)≤u(n−n−k−1))≤(Cu)k+1+Ce−cn.\mathbb{P}\left(s_1(C_k) \le u\left(\sqrt{n}-\sqrt{n-k-1}\right)\right) \le(Cu)^{k+1} + Ce^{-cn}.

Since n−n−k−1≥k/(2n)\sqrt{n}-\sqrt{n-k-1} \ge k/(2\sqrt{n}), the variable Zk=nsk(C)/kZ_k=\sqrt{n}s_k(C)/k therefore satisfies

P(Zk≤u)≤(C′u)k+1+Ce−cn.(34)\mathbb{P}(Z_k \le u) \le(C'u)^{k+1} + Ce^{-cn}. \tag*{(34)}

On QnQ_n it obeys Zk≥an,k:=(log⁡n)−3/kZ_k \ge a_{n,k}:=(\log n)^{-3}/k. Fix u∗>0u_* > 0 such that C′u∗<1/2C'u_* < 1/2. For sufficiently large nn, an,k<u∗a_{n,k} < u_* uniformly for 1≤k≤ℓ1 \le k \le\ell, and tail integration yields

E[1QnZk−p]≤u∗−p+p∫an,ku∗u−p−1(C′u)k+1 du+Ce−cnan,k−p≤Cp.\mathbb{E}\left[\mathbf{1}_{Q_n}Z_k^{-p}\right] \le u_*^{-p} + p\int_{a_{n,k}}^{u_*}u^{-p-1}(C'u)^{k+1}\,du + Ce^{-cn}a_{n,k}^{-p} \le C_p.

Here k+1≥2>pk+1 \ge2 > p, and the last term is uniformly bounded for k≤ℓk \le\ell. This proves (4.6). Summing (4.5) gives Cp∑k≥1k−p<∞C_p\sum_{k\ge1}k^{-p}<\infty. Smaller positive pp 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 ns1(C)\sqrt{n}s_1(C) has quadratic order, while the intensity is at least a constant times (ns1(C))−1(\sqrt{n}s_1(C))^{-1}. The proof uses a fixed p∈(1,2)p \in(1,2) and does not require a second-moment estimate for the intensity.

Lemma 4.3 (Gaussian linearization). For full-rank CC, define

pC(s)=PG0,G(∣G0∣≤sYC),p_C(s)=\mathbb{P}_{G_0,G}(|G_0| \le sY_C),

where G0G_0 is independent of G1,…,GℓG_1,\ldots,G_\ell. On QnQ_n, for s≥0s \ge0,

∣pC(s)−sΛn(C)∣≤Cs3(log⁡n)9ℓ3/2,pC(s)≤sΛn(C).(35)|p_C(s)-s\Lambda_n(C)| \le Cs^3(\log n)^9\ell^{3/2}, \qquad p_C(s) \le s\Lambda_n(C). \tag*{(35)}

Proof. For every x≥0x \ge0, 0≤2/π x−P(∣G0∣≤x)≤Cx30 \le\sqrt{2/\pi}\,x-\mathbb{P}(|G_0| \le x) \le Cx^3. On QnQ_n, YC≤(log⁡n)3∥G∥2Y_C \le(\log n)^3\lVert G\rVert_2, and E∥G∥23≤Cℓ3/2\mathbb{E}\lVert G\rVert_2^3 \le C\ell^{3/2}. Conditioning on YCY_C 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 bb, β\beta, κ\kappa and δn=(log⁡n)−κ\delta_n=(\log n)^{-\kappa}, together with a full event Rn=Rn(C,X)R_n=R_n(C,X) and a deleted-matrix event Qn=Qn(C)Q_n=Q_n(C), with the following properties. The latter event is signed-permutation invariant, contains only full-rank matrices, and implies s1(C)≥n−1/2(log⁡n)−3s_1(C)\ge n^{-1/2}(\log n)^{-3}. Uniformly for e−bn≤s≤n−βe^{-bn}\le s\le n^{-\beta},

P(Ss∩(Rnc∪Qnc))≤Cδns+Ce−bn.(36)\mathbb{P}\left(S_s\cap(R_n^c\cup Q_n^c)\right)\le C\delta_n s+Ce^{-bn}. \tag*{(36)}

On Rn\mathcal{R}_n,

χC(X)≤ΞC(X)≤(1+Cδn)χC(X).(37)\chi_C(X) \le\Xi_C(X) \le(1+C\delta_n)\chi_C(X). \tag*{(37)}

For every fixed C∈QnC \in\mathcal{Q}_n and every such ss,

pC((1−δn)s)−Cδns≤PX(∣v⊤X∣≤sn−1/2χC(X))≤pC((1+δn)s)+Cδns.(38)\begin{aligned} p_C((1-\delta_n)s)-C\delta_n s \le\mathbb{P}_X(|v^\top X|\le sn^{-1/2}\chi_C(X)) \\ &\le p_C((1+\delta_n)s)+C\delta_n s. \tag*{(38)} \end{aligned}

Finally, for the iid square matrix,

P(Ss)=(1+o(1))s+O(e−bn)(39)\mathbb{P}(S_s)=(1+o(1))s+O(e^{-bn}) \tag*{(39)}

uniformly on any sufficiently small subwindow of the displayed range.

Source and conventions. Apply [10] to A⊤A^\top. 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 bb and κ\kappa and increase β\beta if needed, choosing the displayed closed threshold interval strictly inside the source’s open intervals. Take bb smaller also than the exponential constants in (2.5) and (2.7).

The event Qn\mathcal{Q}_n is their E∗\mathcal{E}^* 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 Rn\mathcal{R}_n includes the fresh-vector projection conditions from their Section 4. These conditions depend on XX, so Rn\mathcal{R}_n must be kept distinct from the deleted-matrix event Qn\mathcal{Q}_n.

Theorem 5.2 (Transfer for orbit-bounded marks). Let a=a(C)≥0a=a(C)\ge0 be measurable, set to zero on rank deficiency, and suppose

EP∈Gna(PC)2≤Ka(40)\mathbb{E}_{P\in\mathcal{G}_n}a(PC)^2\le K_a \tag*{(40)}

for every full-rank CC, where Ka<∞K_a<\infty is fixed. Then there exist 0<c0<c10<c_0<c_1, D>1D>1, and a deterministic sequence ηn→0\eta_n\to0, depending only on the entry law and KaK_a, such that

∣P(Tta)−tE[a(C)Λn(C)]∣≤Ctηn+Ce−c1n,e−c0n≤t≤n−D.(41)\left|\mathbb{P}(T_{ta})-t\mathbb{E}[a(C)\Lambda_n(C)]\right|\le Ct\eta_n+Ce^{-c_1n},\qquad e^{-c_0n}\le t\le n^{-D}. \tag*{(41)}

The estimate is uniform over all marks satisfying (5.5). No concentration of aa around one is required.

Proof. We may suppose Ka≥1K_a\ge1. Write δ=δn\delta=\delta_n from Proposition 5.1, choose

L=Ln=min⁡{δ−1/2,log⁡n},h=e−b′n,(42)L=L_n=\min\{\delta^{-1/2},\log n\},\qquad h=e^{-b'n}, \tag*{(42)}

where 0<b′<b0<b'<b is fixed, and put J={h/t≤a≤L}J=\{h/t\le a\le L\}. Choose c0<c1<b′c_0<c_1<b' and D>max⁡{β+1,2}D>\max\{\beta+1,2\}. These choices leave fixed slack for all threshold factors 1+O(δ)1+O(\delta) below. All assertions are for sufficiently large nn.

Removing large and very small marks. Lemma 3.1, applied to TsT_s and (2.5), yields

P(Tta∩{a>L})≤Ct/L+Ce−bn.(43)\mathbb{P}(T_{ta}\cap\{a>L\})\le Ct/L+Ce^{-bn}. \tag*{(43)}

By monotonicity and Lemma 2.2,

P(Tta∩{ta<h})≤P(Th)≤Ch+Ce−bn.(44)\mathbb{P}(T_{ta}\cap\{ta<h\})\le\mathbb{P}(T_h)\le Ch+Ce^{-bn}. \tag*{(44)}

For the coefficient, invariance of Λn\Lambda_n, Lemma 3.1, and Proposition 4.1 give

tE[aΛn1Jc]≤Ct/L+Ch.(45)t\mathbb{E}[a\Lambda_n1_{J^c}] \le Ct/L + Ch. \tag*{(45)}

Indeed the large-mark part is bounded by tKaEΛn/LtK_a\mathbb{E}\Lambda_n/L, and the small-mark part by hEΛnh\mathbb{E}\Lambda_n.

Unmarked regularity at the enlarged threshold. On {a≤L}\{a \le L\}, Tta⊂TtL⊂StLT_{ta} \subset T_{tL} \subset S_{tL}. Consequently,

P(Tta∩J∩(Rnc∪Qnc))≤CδtL+Ce−bn.(46)\mathbb{P}(T_{ta} \cap J \cap(R_n^c \cup Q_n^c)) \le C\delta tL + Ce^{-bn}. \tag*{(46)}

Only the unmarked estimate (36) is used. Since tL≤n−βtL \le n^{-\beta} and tL≥e−bntL \ge e^{-bn}, its range is respected.

Let

qX(s;J)=EC1Qn∩JPX(∣v⊤X∣≤sa(C)n−1/2χC(X)).q_X(s;J) = \mathbb{E}C1_{Q_n\cap J}\mathbb{P}_X(|v^\top X| \le sa(C)n^{-1/2}\chi_C(X)).

The inequality χC≤ΞC\chi_C \le\Xi_C gives qX(t;J)≤P(Tta)q_X(t;J) \le\mathbb{P}(T_{ta}). Conversely, (37), (43)–(46) give, for a fixed constant C⋆>0C_\star> 0,

qX(t;J)≤P(Tta)≤qX((1+C⋆δ)t;J)+Ct(δL+L−1)+Ch+Ce−bn.(47)q_X(t;J) \le\mathbb{P}(T_{ta}) \le q_X((1+C_\star\delta)t;J) + Ct(\delta L + L^{-1}) + Ch + Ce^{-bn}. \tag*{(47)}

Conditional replacement and linearization. For C∈Qn∩JC \in Q_n \cap J, the effective threshold ta(C)ta(C) lies between hh and tLtL. The fixed endpoint slack also permits every threshold αta(C)\alpha ta(C) with 1≤α≤1+C⋆δ1 \le\alpha\le1+C_\star\delta, so (38) applies pointwise throughout this interval. Its averaged error is bounded by CδtEa≤CδtKaC\delta t\mathbb{E}a \le C\delta t\sqrt{K_a}. Lemma 4.3, used at thresholds (1±δ)αta(C)(1\pm\delta)\alpha ta(C), and a31J≤L2aa^3 1_J \le L^2a show that

sup⁡1≤α≤1+C⋆δ∣qX(αt;J)−tE[aΛn1J]∣≤Cδt+Ct3L2(log⁡n)9ℓ3/2.(48)\sup_{1\le\alpha\le1+C_\star\delta}\left|q_X(\alpha t;J)-t\mathbb{E}[a\Lambda_n1_J]\right| \le C\delta t + Ct^3L^2(\log n)^9\ell^{3/2}. \tag*{(48)}

The threshold displacement costs O(δt)O(\delta t) because

E[aΛn]=E[ΛnEPa(PC)]≤KaEΛn≤C.\mathbb{E}[a\Lambda_n] = \mathbb{E}[\Lambda_n\mathbb{E}_P a(PC)] \le\sqrt{K_a}\mathbb{E}\Lambda_n \le C.

Combining (45), (47), and (48) proves (41), for example with

ηn=C(δL+L−1+n−2DL2(log⁡n)9ℓ3/2)→0.\eta_n = C\left(\delta L + L^{-1} + n^{-2D}L^2(\log n)^9\ell^{3/2}\right) \to0.

The terms hh and e−bne^{-bn} are O(e−c1n)O(e^{-c_1n}). None of the errors requires ηn=o(t)\eta_n = o(t).

Remark 5.3 (Why the proof preserves exponential precision). The truncation uses an orbitwise bound inside TtaT_{ta}, rather than the unconditional probability of {a>L}\{a > L\}. Thus its cost is t/Lt/L. The bounded part loses tδnLt\delta_nL. Taking L→∞L \to\infty slowly makes both errors o(t)o(t) even for t=e−αnt=e^{-\alpha n}. 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 EΛn(C)=1+o(1)\mathbb{E}\Lambda_n(C)=1+o(1).

Proof. Choose a fixed D′>DD' > D sufficiently large that t=n−D′t=n^{-D'} 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

P(Tt)=tEΛn+o(t).\mathbb{P}(T_t)=t\mathbb{E}\Lambda_n+o(t).

It remains to compare TtT_t with StS_t. For B=I/nB=I/\sqrt{n}, the rank-one norm in (13) is at least 1/t1/t precisely on TtT_t, and DC≤1/(ns1(C′))D_C \le1/(\sqrt{n}s_1(C')). If smin⁡(A)≤t/((1+γ)n)s_{\min}(A) \le t/((1+\gamma)\sqrt{n}) and DC≤γ/tD_C \le\gamma/t, then (13) implies TtT_t, away from square singularity. Hence

P(St/(1+γ))−Ct2/γ2−Ce−cn≤P(Tt)≤P(St).\mathbb{P}(S_{t/(1+\gamma)})-Ct^2/\gamma^2-Ce^{-cn}\le\mathbb{P}(T_t)\le\mathbb{P}(S_t).

Set γ=t1/3\gamma=t^{1/3} and use (39). Both bounds equal t(1+o(1))t(1+o(1)). 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 B∈Rn×mB \in\mathbb{R}^{n \times m} with ∥B∥HS=1\lVert B\rVert_{\mathrm{HS}}=1, the failure probability

pn(ε,B)=P(∥An−1B∥op≥ε−1)p_n(\varepsilon,B)=\mathbb{P}\left(\lVert A_n^{-1}B\rVert_{\mathrm{op}}\geq\varepsilon^{-1}\right)

satisfies

∣pn(ε,B)−εΘn(B)∣≤Cε[rn+(εn∥B∥op2)1/3]+Ce−c1n(49)\left|p_n(\varepsilon,B)-\varepsilon\Theta_n(B)\right|\leq C\varepsilon\left[r_n+\left(\varepsilon n\lVert B\rVert_{\mathrm{op}}^2\right)^{1/3}\right]+Ce^{-c_1n} \tag*{(49)}

in a window (1.4), uniformly in mm and BB. The inverse norm is assigned +∞+\infty on square singularity.

Proof. The mark aBa_B satisfies (5.5) with Ka=1K_a=1. By Theorem 5.2,

P(TtaB)=tE[aBΛn]+o(t)+O(e−cn)\mathbb{P}(T_{ta_B})=t\mathbb{E}[a_B\Lambda_n]+o(t)+O(e^{-cn})

uniformly in BB 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 L4/3L^{4/3} norm by Proposition 4.1. Therefore Theorem 3.3 and Proposition 5.4 give

E[aBΛn]=Θn(B)+o(1)(50)\mathbb{E}[a_B\Lambda_n]=\Theta_n(B)+o(1) \tag*{(50)}

uniformly over mm, BB. Its boundedness also follows directly from orbit averaging. Apply (2.6) at thresholds ε/(1±δ)\varepsilon/(1\pm\delta). Choose the main window strictly inside the transfer window. The threshold displacements cost O(εδ)O(\varepsilon\delta), while (2.7) costs O(ε2n∥B∥op2/δ2)+O(e−cn)O(\varepsilon^2n\lVert B\rVert_{\mathrm{op}}^2/\delta^2)+O(e^{-cn}). Setting

δ=(εn∥B∥op2)1/3\delta=\left(\varepsilon n\lVert B\rVert_{\mathrm{op}}^2\right)^{1/3}

balances these errors. It is less than 1/41/4 uniformly for large nn once C0>1C_0>1. Gather the uniform vanishing coefficient and transfer errors into rnr_n. Square and deleted singularity probabilities are absorbed into the exponential remainder, decreasing c1c_1 and then c0c_0 if necessary so that c0<c1c_0<c_1. This proves (6.1). □

Proof of Theorem 1.1. For nonsingular AnA_n, with BB as in (1.2),

smin⁡(MAn)≤ε/∥M−1∥HS⟺∥An−1B∥op≥ε−1.s_{\min}(MA_n)\leq\varepsilon/\lVert M^{-1}\rVert_{\mathrm{HS}} \quad\Longleftrightarrow\quad \lVert A_n^{-1}B\rVert_{\mathrm{op}}\geq\varepsilon^{-1}.

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 H⪰0H\succeq0 with Tr⁡H=1\operatorname{Tr}H=1. For the functional ϑn\vartheta_n defined in (1.3),

2π≤n E∣U1∣≤ϑn(H)≤1,(51)\sqrt{\frac{2}{\pi}}\leq\sqrt{n}\,\mathbb{E}|U_1|\leq\vartheta_n(H)\leq1, \tag*{(51)}

and

0≤1−ϑn(H)≤nn+2(Tr⁡H2−1n).(52)0\leq1-\vartheta_n(H)\leq\frac{n}{n+2}\left(\operatorname{Tr}H^2-\frac{1}{n}\right). \tag*{(52)}

For a sequence HnH_n, ϑn(Hn)→1\vartheta_n(H_n)\to1 if and only if Tr⁡Hn2→0\operatorname{Tr}H_n^2\to0.

Proof. The upper bound is Jensen’s inequality. If λj\lambda_j are the eigenvalues of HH, concavity of the square root gives pointwise

∑jλjnUj2≥∑jλjn∣Uj∣.\sqrt{\sum_j \lambda_j nU_j^2} \ge\sum_j \lambda_j \sqrt{n}|U_j|.

Taking expectations gives the middle lower bound. If g=RUg=RU is the Gaussian polar decomposition, RR and UU are independent and ER≤n\mathbb{E}R\le\sqrt{n}. Thus nE∣U1∣=nE∣g1∣/ER≥2/π\sqrt{n}\mathbb{E}|U_1|=\sqrt{n}\mathbb{E}|g_1|/\mathbb{E}R\ge\sqrt{2/\pi}.

Set S=nUTHUS=nU^\mathsf{T}HU. The sphere moment formulas give

ES=1,Var⁡(S)=2nn+2(Tr⁡H2−1n).\mathbb{E}S=1,\qquad\operatorname{Var}(S)=\frac{2n}{n+2}\left(\operatorname{Tr}H^2-\frac{1}{n}\right).

The exact identity 1−ES=12E(S−1)21-\mathbb{E}\sqrt{S}=\frac{1}{2}\mathbb{E}(\sqrt{S}-1)^2, together with (S−1)2≤(S−1)2(\sqrt{S}-1)^2\le(S-1)^2, proves (7.2).

For necessity, ϑn(Hn)→1\vartheta_n(H_n)\to1 implies S→1\sqrt{S}\to1 in L2L^2, hence S→1S\to1 in probability. By convexity,

ES4≤∑jλjE(nUj2)4=105n4n(n+2)(n+4)(n+6)≤105.\mathbb{E}S^4\le\sum_j\lambda_j\mathbb{E}(nU_j^2)^4=\frac{105n^4}{n(n+2)(n+4)(n+6)}\le105.

Thus S2S^2 is uniformly integrable and Var⁡(S)→0\operatorname{Var}(S)\to0. The variance formula implies Tr⁡Hn2→0\operatorname{Tr}H_n^2\to0. Sufficiency follows from (7.2).

Proof of Corollary 1.3. Combine (7.2) with Theorem 1.1.

Remark 7.2 (Effective ranks). For Tr⁡H=1\operatorname{Tr}H=1 and H=BBTH=BB^\mathsf{T},

∥B∥op4≤Tr⁡H2≤∥B∥op2.\|B\|_{\mathrm{op}}^4\le\operatorname{Tr}H^2\le\|B\|_{\mathrm{op}}^2.

Thus divergence of 1/Tr⁡H21/\operatorname{Tr}H^2 and divergence of the operator stable rank 1/∥B∥op21/\|B\|_{\mathrm{op}}^2 are equivalent qualitative conditions. They encode different quantitative information. The coefficient bias in (7.2) is quadratic in Δ4\Delta_4, whereas the inverse residual is governed by ∥B∥op\|B\|_{\mathrm{op}}.

Proposition 7.3 (General spectral profiles). Let the eigenvalues of HnH_n be λn,1≥⋯≥λn,n≥0\lambda_{n,1}\ge\cdots\ge\lambda_{n,n}\ge0, with sum one. Suppose λn,j→λj\lambda_{n,j}\to\lambda_j for every fixed jj, and put d=1−∑j≥1λj≥0d=1-\sum_{j\ge1}\lambda_j\ge0. Then

ϑn(Hn)⟶Ed+∑j≥1λjgj2.(53)\vartheta_n(H_n)\longrightarrow\mathbb{E}\sqrt{d+\sum_{j\ge1}\lambda_jg_j^2}. \tag*{(53)}

The series is almost surely finite. In particular, for rr equal surviving weights carrying all the mass, the limit is

Eχr2/r=2rΓ((r+1)/2)Γ(r/2).(54)\mathbb{E}\sqrt{\chi_r^2/r}=\sqrt{\frac{2}{r}}\frac{\Gamma((r+1)/2)}{\Gamma(r/2)}. \tag*{(54)}

Proof. Gaussian polar decomposition gives the exact identity

ϑn(Hn)=nE∥g(n)∥2E∑j=1nλn,jgj2.(55)\vartheta_n(H_n)=\frac{\sqrt{n}}{\mathbb{E}\|g^{(n)}\|_2}\mathbb{E}\sqrt{\sum_{j=1}^{n}\lambda_{n,j}g_j^2}. \tag*{(55)}

The prefactor tends to one. For fixed rr, the first rr weighted terms converge in distribution. The centered remainder has variance

2∑j>rλn,j2≤2λn,r+1.2\sum_{j>r}\lambda_{n,j}^2\le2\lambda_{n,r+1}.

Its mean tends to 1−∑j≤rλj1-\sum_{j\le r}\lambda_j. Taking n→∞n\to\infty first and then r→∞r\to\infty identifies the limit of the quadratic sum as d+∑j≥1λjgj2d+\sum_{j\ge1}\lambda_jg_j^2, because λr+1→0\lambda_{r+1}\to0. 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 Θn(B)=ϑn(BB⊤)\Theta_n(B)=\vartheta_n(BB^\top) 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 Kn∈Rn×mnK_n\in\mathbb{R}^{n\times m_n}, uniformly in (1.4),

P(∥An−1Kn∥op≥∥Kn∥HSε)=ε[Θn(Kn∥Kn∥HS)+o(1)]+O(e−c1n).(56)\mathbb{P}\left(\left\lVert A_n^{-1}K_n\right\rVert_{\mathrm{op}}\ge\frac{\lVert K_n\rVert_{\mathrm{HS}}}{\varepsilon}\right)=\varepsilon\left[\Theta_n\left(\frac{K_n}{\lVert K_n\rVert_{\mathrm{HS}}}\right)+o(1)\right]+O\left(e^{-c_1n}\right). \tag*{(56)}

For one nonzero deterministic right-hand side bnb_n,

P(∥An−1bn∥2≥∥bn∥2ε)=(2π+o(1))ε+O(e−c1n).(57)\mathbb{P}\left(\left\lVert A_n^{-1}b_n\right\rVert_2\ge\frac{\lVert b_n\rVert_2}{\varepsilon}\right)=\left(\sqrt{\frac{2}{\pi}}+o(1)\right)\varepsilon+O\left(e^{-c_1n}\right). \tag*{(57)}

Proof. Normalize KnK_n 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 KnB2mnK_nB_2^{m_n}. Taking Kn=IK_n=I gives worst-case amplification, whereas choosing KnK_n 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 Σn\Sigma_n be deterministic positive definite and set

Sn=1nΣn1/2AnAn⊤Σn1/2,Bn=Σn−1/2Tr⁡Σn−1.S_n=\frac{1}{n}\Sigma_n^{1/2}A_nA_n^\top\Sigma_n^{1/2},\qquad B_n=\frac{\Sigma_n^{-1/2}}{\sqrt{\operatorname{Tr}\Sigma_n^{-1}}}.

Then uniformly in (1.4),

P(λmin⁡(Sn)≤ε2nTr⁡Σn−1)=ε[Θn(Bn)+o(1)]+O(e−c1n).(58)\mathbb{P}\left(\lambda_{\min}(S_n)\le\frac{\varepsilon^2}{n\operatorname{Tr}\Sigma_n^{-1}}\right)=\varepsilon\left[\Theta_n(B_n)+o(1)\right]+O\left(e^{-c_1n}\right). \tag*{(58)}

Proof. Use λmin⁡(Sn)=n−1smin⁡(Σn1/2An)2\lambda_{\min}(S_n)=n^{-1}s_{\min}(\Sigma_n^{1/2}A_n)^2 and apply Theorem 1.1.

The conclusion holds for arbitrary positive-definite Σn\Sigma_n, with symmetric iid latent entries and thresholds in the stated shrinking window.

Independent random environments

Corollary 8.3 (Conditional universality). Let En\mathcal{E}_n be a sigma-field independent of AnA_n, and let MnM_n be an almost surely invertible En\mathcal{E}_n-measurable matrix. Put Bn=Mn−1/∥Mn−1∥HSB_n=M_n^{-1}/\lVert M_n^{-1}\rVert_{\mathrm{HS}}. For deterministic ε\varepsilon in (1.4), almost surely,

P(smin⁡(MnAn)≤ε∥Mn−1∥HS∣En)=ε[Θn(Bn)+O(r~n)]+O(e−c1n).(59)\mathbb{P}\left(s_{\min}(M_nA_n)\le\left.\frac{\varepsilon}{\lVert M_n^{-1}\rVert_{\mathrm{HS}}}\right|\mathcal{E}_n\right)=\varepsilon\left[\Theta_n(B_n)+O(\widetilde r_n)\right]+O\left(e^{-c_1n}\right). \tag*{(59)}

where r~n→0\widetilde r_n\to0 and all error constants are deterministic and uniform in the environment.

Proof. Condition on En\mathcal{E}_n and use the uniform bound of Theorem 1.1. The residual is bounded by (εn)1/3(\varepsilon_n)^{1/3} and can be included in r~n\widetilde r_n.

For two independent fresh iid matrices, also independent of the same environment, let En,1,En,2E_{n,1}, E_{n,2} be the corresponding failures with the same MnM_n and threshold. Conditional independence and (59) imply

ε−2Cov⁡(1En,1,1En,2)=Var⁡(Θn(Bn))+o(1).(60)\varepsilon^{-2}\operatorname{Cov}(\mathbf{1}_{E_{n,1}},\mathbf{1}_{E_{n,2}})=\operatorname{Var}(\Theta_n(B_n))+o(1). \tag*{(60)}

Indeed the covariance equals the variance of the conditional failure probability, and its ratio to ε\varepsilon differs uniformly from Θn(Bn)\Theta_n(B_n) by o(1)o(1). Thus concentration of the coefficient is sufficient for asymptotic decorrelation in this conditional-independence setting.

Independent exponential families

Corollary 8.4 (Independent extremes). Let Nn=ehn+o(n)N_n=e^{hn+o(n)} with 0<h<min⁡{c0,c1}0<h<\min\{c_0,c_1\}. Let An,jA_{n,j}, 1≤j≤Nn1\le j\le N_n, be independent copies of AnA_n, and let Mn,jM_{n,j} be deterministic invertible matrices. Define

Zn,j=∥Mn,j−1∥HSsmin⁡(Mn,jAn,j),θn,j=Θn(Mn,j−1∥Mn,j−1∥HS).Z_{n,j}=\lVert M_{n,j}^{-1}\rVert_{\mathrm{HS}}s_{\min}(M_{n,j}A_{n,j}),\qquad\theta_{n,j}=\Theta_n\left(\frac{M_{n,j}^{-1}}{\lVert M_{n,j}^{-1}\rVert_{\mathrm{HS}}}\right).

If Nn−1∑j≤Nnθn,j→θˉN_n^{-1}\sum_{j\le N_n}\theta_{n,j}\to\bar\theta, then for each z>0z>0,

∑j=1Nn1{Zn,j≤z/Nn}⟹Poisson⁡(zθˉ),(61)\sum_{j=1}^{N_n}\mathbf{1}_{\{Z_{n,j}\le z/N_n\}}\Longrightarrow\operatorname{Poisson}(z\bar\theta), \tag*{(61)}

and Nnmin⁡j≤NnZn,jN_n\min_{j\le N_n}Z_{n,j} converges in distribution to an exponential random variable of rate θˉ\bar\theta.

Proof. The threshold z/Nnz/N_n lies in (4) for large nn. Theorem 1.1, uniformly in jj, gives probabilities pn,jp_{n,j} with ∑jpn,j→zθˉ\sum_j p_{n,j}\to z\bar\theta and max⁡jpn,j→0\max_jp_{n,j}\to0. The accumulated exponential floor tends to zero because Nne−c1n→0N_ne^{-c_1n}\to0. 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 LpL^p 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 v=v(C)v=v(C) and δn↓0\delta_n\downarrow0, the random positive-definite matrix

H(C)=I−vvT+δnvvTn−1+δnH(C)=\frac{I-vv^{\mathsf T}+\delta_nvv^{\mathsf T}}{n-1+\delta_n}

has trace one and diffuse squared spectral mass, but nvTH(C)v=nδn/(n−1+δn)→0nv^{\mathsf T}H(C)v=n\delta_n/(n-1+\delta_n)\to0. 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 q≥1q \ge1. Let An,q+A^+_{n,q} be an n×(n+q−1)n \times(n+q-1) matrix with iid entries of law ξ\xi, and let Gn,q+G^+_{n,q} have iid standard real Gaussian entries. Here smin⁡s_{\min} denotes the least of the nn singular values, and equals zero on rank deficiency. For M∈GLn(R)M \in GL_n(\mathbb{R}), retain the normalization h(M)h(M) from (1.2) and write

Fn,q,Mξ(t)=P(h(M)smin⁡(MAn,q+)≤t),F^{\xi}_{n,q,M}(t)=\mathbb{P}\bigl(h(M)s_{\min}(MA^+_{n,q})\le t\bigr),

and define Fn,q,MGF^G_{n,q,M} analogously.

Conjecture 10.1 (Uniform Gaussian comparison across hard-edge scales). Let ξ\xi be a fixed real, centered, variance-one, subgaussian random variable. Symmetry is not assumed. For each fixed q≥1q \ge1, there exists cξ,q>0c_{\xi,q}>0 such that, for every fixed T>0T>0,

sup⁡M∈GLn(R)sup⁡e−cξ,qn≤t≤T∣Fn,q,Mξ(t)Fn,q,MG(t)−1∣⟶0.(62)\sup_{M\in GL_n(\mathbb{R})}\sup_{e^{-c_{\xi,q}n}\le t\le T}\left|\frac{F^{\xi}_{n,q,M}(t)}{F^G_{n,q,M}(t)}-1\right|\longrightarrow0. \tag*{(62)}

For M=IM=I and q=1q=1, 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 MM, 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 o(1)o(1) 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 M=IM=I, its limiting distribution is 1−e−t−t2/21-e^{-t-t^2/2} [1, 11]. For a rank-one input, the reciprocal Gaussian amplification is exactly ∣g∣|g|. 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 Θn\Theta_n in Theorem 1.1 is to describe the first-order rare-tail behavior.

The exponential endpoint must depend on the entry law and on qq. For example, for ξ=(Bernoulli⁡(p)−p)/p(1−p)\xi= (\operatorname{Bernoulli}(p)-p)/\sqrt{p(1-p)}, coincidence of two rows of An,q+A^+_{n,q} has probability [p2+(1−p)2]n+q−1[p^2+(1-p)^2]^{n+q-1} 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

1q!Ψn,q(H),Ψn,q(H)=EU∼Haar⁡(Sn−1)(nUTHU)q/2,H=M−1M−Th(M)2.(63)\frac{1}{q!}\Psi_{n,q}(H), \qquad\Psi_{n,q}(H)=\mathbb{E}_{U\sim\operatorname{Haar}(S^{n-1})}(nU^\mathsf{T}HU)^{q/2}, \qquad H=\frac{M^{-1}M^{-\mathsf{T}}}{h(M)^2}. \tag*{(63)}

The proposed asymptotic is a lower tail of order tqt^q with coefficient Ψn,q(H)/q!\Psi_{n,q}(H)/q! as n→∞n\to\infty. Exact fixed-dimensional Gaussian coefficients may include further normalization factors. At q=1q=1 it is the coefficient of this paper. At q=2q=2, Ψn,2=1\Psi_{n,2}=1 exactly. For q>2q>2, the functional is convex and symmetric in the eigenvalues of HH, so concentrating inverse singular mass increases it in the majorization order, reversing the concave ordering at q=1q=1. A rank-one profile gives E∣g∣q/q!=1/[2q/2Γ(q/2+1)]\mathbb{E}|g|^q/q!=1/[2^{q/2}\Gamma(q/2+1)] in the large-nn limit, matching the lower-tail coefficient of χq\chi_q.

A rectangular intrinsic-volume hierarchy

Now let An,q−A^-_{n,q} have size n×(n−q+1)n\times(n-q+1) and iid entries of law ξ\xi, where q≥1q\ge1 is fixed and n≥qn\ge q. The multiplier still acts on the nn rows. For trace-one H⪰0H\succeq0, define

Φn,q(H)=EQdet⁡(nQTHQ),QTQ=Iq,(64)\Phi_{n,q}(H)=\mathbb{E}_Q\sqrt{\det(nQ^\mathsf{T}HQ)}, \qquad Q^\mathsf{T}Q=I_q, \tag*{(64)}

where the columns of QQ form a Haar-uniform orthonormal qq-frame in Rn\mathbb{R}^n. Thus Φn,1(H)=ϑn(H)\Phi_{n,1}(H)=\vartheta_n(H).

Conjecture 10.2 (Intrinsic-volume coefficients at a rectangular hard edge). Assume (1.1). For each fixed q≥1q\ge1, there are constants c0,c1,C0,C>0c_0,c_1,C_0,C>0, with qc0<c1qc_0<c_1, and a deterministic rn,q→0r_{n,q}\to0, depending only on qq and the entry law, such that, uniformly over M∈GLn(R)M\in GL_n(\mathbb{R}) and e−c0n≤ε≤n−C0e^{-c_0n}\le\varepsilon\le n^{-C_0}

∣P(smin⁡(MAn,q−)≤εh(M))−εqq!Φn,q(H)∣≤rn,qεq+Ce−c1n,(65)\left|\mathbb{P}\left(s_{\min}(MA^-_{n,q})\le\frac{\varepsilon}{h(M)}\right)-\frac{\varepsilon^q}{q!}\Phi_{n,q}(H)\right|\le r_{n,q}\varepsilon^q+Ce^{-c_1n}, \tag*{(65)}

where H=M−1M−T/h(M)2H=M^{-1}M^{-\mathsf{T}}/h(M)^2 and smin⁡s_{\min} is zero on rank deficiency.

The coefficients have an intrinsic-volume interpretation. Let κj\kappa_j be the volume of the unit ball in Rj\mathbb{R}^j, and let VqV_q denote the intrinsic volume with the Steiner normalization. Kubota’s formula gives

Φn,q(H)=nq/2κn−q(nq)κnVq(H1/2Bn).(66)\Phi_{n,q}(H)=\frac{n^{q/2}\kappa_{n-q}}{\binom{n}{q}\kappa_n}V_q(H^{1/2}B^n). \tag*{(66)}

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 0≤Φn,q≤10\le\Phi_{n,q}\le1, but for q≥2q\ge2 this coefficient can tend to zero. Indeed, Cauchy–Binet and Haar symmetry give

EQdet⁡(nQTHQ)=nqeq(H)(nq)≤1,\mathbb{E}_Q\det(nQ^\mathsf{T}HQ)=\frac{n^q e_q(H)}{\binom{n}{q}}\le1,

where eqe_q is the qqth elementary symmetric polynomial. Jensen’s inequality then gives the upper bound on Φn,q\Phi_{n,q}. If all inverse mass concentrates on r<qr<q limiting directions, the projected determinant vanishes. Under a finite-spike profile with weights λ1,…,λr\lambda_1,\ldots,\lambda_r and diffuse residual mass dd, the corresponding geometric limit is

Φn,q(Hn)⟶Edet⁡(dIq+∑j=1rλjgjgjT),gj∼N(0,Iq) independently.(67)\Phi_{n,q}(H_n) \longrightarrow\mathbb{E}\sqrt{\det\left(dI_q+\sum_{j=1}^{r}\lambda_j g_jg_j^{\mathsf{T}}\right)},\qquad g_j\sim N(0,I_q)\text{ independently}. \tag*{(67)}

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 Φn,q\Phi_{n,q} 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 Ma=diag⁡(a,1,…,1)M_a=\operatorname{diag}(a,1,\ldots,1) and q≥2q\ge2, the Gaussian lower (n−1)×(n−q+1)(n-1)\times(n-q+1) block has full column rank almost surely, while ε/h(Ma)→0\varepsilon/h(M_a)\to0 as a↓0a\downarrow0. 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 Gn,q+G_{n,q}^{+} be as above. If V∈Rn×rV\in\mathbb{R}^{n\times r} has orthonormal columns and DD is a positive diagonal r×rr\times r matrix, then

∥(Gn,q+)†VD∥op=d∥(Gr,q+)†D∥op.(68)\left\|\left(G_{n,q}^{+}\right)^{\dagger}VD\right\|_{\mathrm{op}} \overset{d}{=} \left\|\left(G_{r,q}^{+}\right)^{\dagger}D\right\|_{\mathrm{op}}. \tag*{(68)}

Here †\dagger denotes the Moore–Penrose inverse. In particular, the reciprocal for r=1r=1, D=1D=1 has law χq\chi_q. Moreover, for q=2q=2, every nn, every invertible MM, and every t≥0t\ge0,

Fn,2,MG(t)=1−e−t2/2.(69)F_{n,2,M}^{G}(t)=1-e^{-t^2/2}. \tag*{(69)}

Proof. Put W=Gn,q+(Gn,q+)TW=G_{n,q}^{+}(G_{n,q}^{+})^{\mathsf{T}}. The squared norm in (10.7) equals

λmax⁡(DVTW−1VD).\lambda_{\max}\left(DV^{\mathsf{T}}W^{-1}VD\right).

Rotate VV to the first rr coordinate vectors and condition on the remaining n−rn-r Gaussian rows. Their orthogonal complement has dimension r+q−1r+q-1, and the projected first rr rows form an iid r×(r+q−1)r\times(r+q-1) Gaussian matrix. The Schur complement identifies the leading r×rr\times r block of W−1W^{-1} with the inverse Gram matrix of this projection, proving (10.7).

For (10.8), W=Gn,2+(Gn,2+)TW=G_{n,2}^{+}(G_{n,2}^{+})^{\mathsf{T}} has density proportional to e−Tr⁡W/2e^{-\operatorname{Tr}W/2} on the positive-definite cone. The event in the survival probability is W>t2HW>t^2H. Translation by t2Ht^2H multiplies its integral by e−t2Tr⁡H/2=e−t2/2e^{-t^2\operatorname{Tr}H/2}=e^{-t^2/2}. This is also a special case of the correlated Gaussian Wishart formulas in [13].

Proposition 10.4 (The tall Gaussian coefficient). Fix n≥2n\ge2 and M∈GLn(R)M\in GL_n(\mathbb{R}). Let GG be an n×(n−1)n\times(n-1) matrix with iid standard real Gaussian entries and zz uniform on Sn−1S^{n-1}. With h=h(M)h=h(M) and H=M−1M−T/h2H=M^{-1}M^{-\mathsf{T}}/h^2,

lim⁡ε↓01ε2P(smin⁡(MG)≤εh)=12(1−EzzTH2zzTHz)\lim_{\varepsilon\downarrow0}\frac{1}{\varepsilon^2}\mathbb{P}\left(s_{\min}(MG)\le\frac{\varepsilon}{h}\right) =\frac{1}{2}\left(1-\mathbb{E}_z\frac{z^{\mathsf{T}}H^2z}{z^{\mathsf{T}}Hz}\right)
=n−12nΦn,2(H).(70)=\frac{n-1}{2n}\Phi_{n,2}(H). \tag*{(70)}

Proof. Replacing MM by hMhM, we may assume h=1h=1. Put p=n−1p=n-1. The joint density of the ordered squared singular values λ1<⋯<λp\lambda_1<\cdots<\lambda_p of GG is proportional to

e−∑i=1pλi/2∏i<j(λj−λi).e^{-\sum_{i=1}^{p}\lambda_i/2}\prod_{i<j}(\lambda_j-\lambda_i).

The substitution X=λ1X=\lambda_1 and Yi=λi−λ1Y_i=\lambda_i-\lambda_1, 2≤i≤p2\le i\le p, shows that XX is exponential with rate p/2p/2, independently of (Y2,…,Yp)(Y_2,\ldots,Y_p). Choose the left singular frame UU with independent uniform column signs. It is Haar and independent of these variables. Set

D=diag⁡(0,Y2,…,Yp),K=UTMTMU.D=\operatorname{diag}(0,Y_2,\ldots,Y_p),\qquad K=U^{\mathsf T}M^{\mathsf T}MU.

For p=1p=1, take D=(0)D=(0). The event smin⁡(MG)2>ε2s_{\min}(MG)^2>\varepsilon^2 is equivalent to XIp+D>ε2K−1XI_p+D>\varepsilon^2K^{-1}, and hence to X>λmax⁡(ε2K−1−D)X>\lambda_{\max}(\varepsilon^2K^{-1}-D). Thus

P(smin⁡(MG)≤ε∣U,D)=1−exp⁡[−p2λmax⁡(ε2K−1−D)].\mathbb{P}(s_{\min}(MG)\le\varepsilon\mid U,D)=1-\exp\left[-\frac{p}{2}\lambda_{\max}(\varepsilon^2K^{-1}-D)\right].

The largest eigenvalue on the right is positive, since its Rayleigh quotient at the first coordinate vector is ε2(K−1)11\varepsilon^2(K^{-1})_{11}. The zero eigenvalue of −D-D is simple almost surely, so

ε−2λmax⁡(ε2K−1−D)⟶(K−1)11.\varepsilon^{-2}\lambda_{\max}(\varepsilon^2K^{-1}-D)\longrightarrow(K^{-1})_{11}.

If zz is a unit normal to the columns of UU, the Schur complement yields

K−1=UT(H−HzzTHzTHz)U.K^{-1}=U^{\mathsf T}\left(H-\frac{Hzz^{\mathsf T}H}{z^{\mathsf T}Hz}\right)U.

In particular, ∥K−1∥op≤∥H∥op\lVert K^{-1}\rVert_{\mathrm{op}}\le\lVert H\rVert_{\mathrm{op}}, and the conditional probability divided by ε2\varepsilon^2 is bounded by p∥H∥op/2p\lVert H\rVert_{\mathrm{op}}/2. Dominated convergence therefore gives the coefficient pE(K−1)11/2p\mathbb{E}(K^{-1})_{11}/2. Conditional on zz, the first column of UU is uniform on the unit sphere of z⊥z^\perp. Averaging its quadratic form gives the first equality in (70).

For the second equality, the ellipsoid formula in [2 Section 2] gives

V2(H1/2Bn2)=π(Tr⁡H−EzzTH2zzTHz).V_2(H^{1/2}\mathbb{B}_n^2)=\pi\left(\operatorname{Tr}H-\mathbb{E}_z\frac{z^{\mathsf T}H^2z}{z^{\mathsf T}Hz}\right).

Now apply (66) with q=2q=2, using Tr⁡H=1\operatorname{Tr}H=1 and κn=2πκn−2/n\kappa_n=2\pi\kappa_{n-2}/n.

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 q>1q>1, the rarity scale is εq\varepsilon^q, 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 qq-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 L4L^4 Wasserstein bound for μνn\mu_{\nu_n} with the uniform L4/3L^{4/3} bound for Λn\Lambda_n.

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 n≥4n \ge4, 2≤m≤n2 \le m \le n, and z1,…,zn∈[0,L]z_1,\ldots,z_n \in[0,L] with L≥1L \ge1 and ∑izi=n\sum_i z_i = n. Let II be a uniformly chosen mm-element subset of {1,…,n}\{1,\ldots,n\} containing 1 and 2, and set

S=1m∑i∈Izi.S = \frac{1}{m}\sum_{i\in I} z_i.

Then

∣EIS−1∣≤4Lm,Var⁡I(S)≤2Lm.(71)\left|\mathbb{E}_I S - 1\right| \le\frac{4L}{m}, \qquad\operatorname{Var}_I(S) \le\frac{2L}{m}. \tag*{(71)}

In particular, S→1S \to1 in probability whenever L/m→0L/m \to0.

Proof. Put a0=z1+z2a_0 = z_1 + z_2. The remaining m−2m-2 entries are sampled without replacement from a population of size n−2n-2, and

EI(mS)−m=n−mn−2(a0−2).\mathbb{E}_I(mS)-m = \frac{n-m}{n-2}(a_0-2).

This gives the first bound. The finite-population variance formula, followed by zi2≤Lziz_i^2 \le Lz_i, gives

Var⁡I(mS)≤m−2n−2∑i=3nzi2≤mnLn−2≤2mL.\operatorname{Var}_I(mS) \le\frac{m-2}{n-2}\sum_{i=3}^{n} z_i^2 \le\frac{mnL}{n-2} \le2mL.

The conclusion follows from these bounds and Chebyshev’s inequality.

Proof of Lemma 3.4. Write A=CnT\mathcal{A} = C_n^{\mathsf{T}}, so A\mathcal{A} is an (n−1)×n(n-1)\times n iid matrix, and let vv 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 CC such that

P(max⁡1≤i≤nn∣vi∣2>Clog⁡n)=o(1).(72)\mathbb{P}\left(\max_{1\le i\le n} n|v_i|^2 > C\log n\right) = o(1). \tag*{(72)}

Sampling and normalization. Independently of A\mathcal{A}, choose a uniform mm-element set II containing 1 and 2, ordered with these indices first. The parameter mm will be a sufficiently small positive power of nn. On the event in (A.2) with the reverse inequality, apply Lemma A.1 conditionally to zi=n∣vi∣2z_i = n|v_i|^2 and L=Clog⁡nL = C\log n. Since L/m→0L/m \to0, it follows that

Sn:=nm∑i∈I∣vi∣2⟶1S_n := \frac{n}{m}\sum_{i\in I}|v_i|^2 \longrightarrow1

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 SnS_n.

Projection and Gaussian comparison. Let HH be the span of the columns of A\mathcal{A} indexed by IcI^c. With exponentially high probability, those columns have full column rank, so dim⁡H⊥=m−1\dim H^\perp= m-1. The flatness estimate [6 Lemma 6.3] gives a constant c>0c>0 such that, for sufficiently small polynomial choices of mm, with probability 1−o(1)1-o(1) over II and the omitted columns,

sup⁡u∈H⊥,∥u∥2=1∥u∥∞≤n−c.(73)\sup_{\substack{u\in H^\perp,\,\lVert u\rVert_2=1}} \lVert u\rVert_\infty\le n^{-c}. \tag*{(73)}

Choose a fixed α>0\alpha>0 with α≤min⁡{c/100,1/100}\alpha\le\min\{c/100,1/100\} small enough to meet that estimate, and put

m=⌊nα⌋,ρ=n−c/4.m=\lfloor n^\alpha\rfloor,\qquad\rho=n^{-c/4}.

The constant cc in the flatness estimate can be taken independently of this sufficiently small choice of α\alpha.

Condition only on II and an omitted-column realization satisfying full column rank and (A.4). Choose an orthonormal basis of H⊥H^\perp, and let π:Rn−1→Rm−1\pi:\mathbb{R}^{n-1}\to\mathbb{R}^{m-1} be the corresponding coordinate projection. Thus ππ⊤=Im−1\pi\pi^\top=I_{m-1}. Set Z=πAIZ=\pi\mathcal{A}_I, an (m−1)×m(m-1)\times m matrix. The entries of AI\mathcal{A}_I are still independent copies of the original entry variable under this conditioning.

For 1≤r≤n−11\le r\le n-1 and 1≤j≤m1\le j\le m, let wrjw_{rj} be the vectorization of the matrix (πer)ej⊤(\pi e_r)e_j^\top. Then

vec⁡(Z)=∑r,jξrjwrj,∑r,jwrjwrj⊤=Im(m−1),max⁡r,j∥wrj∥∞≤n−c.\operatorname{vec}(Z)=\sum_{r,j}\xi_{rj}w_{rj},\qquad\sum_{r,j}w_{rj}w_{rj}^{\top}=I_{m(m-1)},\qquad\max_{r,j}\lVert w_{rj}\rVert_\infty\le n^{-c}.

The Gaussian replacement of this normalized frame is therefore the vectorization of an iid standard Gaussian matrix GG of the same dimensions. For any measurable set Ω\Omega in the matrix space, the frame Berry–Esseen estimate [6 Lemma 2.5] gives

PG(G∈Ω∖∂ρΩ)−rnBE≤P(Z∈Ω∣I,AIc)≤PG(G∈Ω∪∂ρΩ)+rnBE.(74)\mathbb{P}_{G}(G\in\Omega\setminus\partial_\rho\Omega)-r_n^{\mathrm{BE}} \le\mathbb{P}(Z\in\Omega\mid I,\mathcal{A}_{I^c}) \le\mathbb{P}_{G}(G\in\Omega\cup\partial_\rho\Omega)+r_n^{\mathrm{BE}}. \tag*{(74)}

where ∂ρΩ\partial_\rho\Omega is the ρ\rho-neighborhood of the boundary in the entrywise maximum norm, and

rnBE≤Cm5ρ−3n−c≤Cn−c/5=o(1).r_n^{\mathrm{BE}}\le Cm^5\rho^{-3}n^{-c}\le Cn^{-c/5}=o(1).

The constants are uniform over the conditioned realizations.

Control of the event boundary. For fixed t1,t2≥0t_1,t_2\ge0, let Ω\Omega consist of full-row-rank matrices whose unit kernel vector yy, defined up to sign, satisfies m∣yj∣≤tj\sqrt{m}|y_j|\le t_j for j=1,2j=1,2. The Gaussian rectangular bound gives

PG(smin⁡(G)<m−2)≤Cm−3+Ce−c′m.(75)\mathbb{P}_{G}(s_{\min}(G)<m^{-2})\le Cm^{-3}+Ce^{-c'm}. \tag*{(75)}

where smin⁡(G)s_{\min}(G) denotes its smallest nonzero singular value on full row rank. Suppose smin⁡(G)≥m−2s_{\min}(G)\ge m^{-2} and G′G' differs from GG by at most ρ\rho in each entry. Then ∥G′−G∥op≤mρ=o(m−2)\lVert G'-G\rVert_{\mathrm{op}}\le m\rho=o(m^{-2}), so G′G' has full row rank. If y′y' is its unit kernel vector, projection onto the row space of GG gives

∥Proj⁡row⁡(G)y′∥2≤∥Gy′∥2smin⁡(G)≤m3ρ.\left\lVert\operatorname{Proj}_{\operatorname{row}(G)}y'\right\rVert_2 \le\frac{\lVert Gy'\rVert_2}{s_{\min}(G)} \le m^3\rho.

After choosing the sign of a unit vector yG∈ker⁡Gy_G\in\ker G, we obtain

∥y′−yG∥2≤Cm3ρ,∣m∣yj′∣−m∣(yG)j∣∣≤Cm7/2ρ,j=1,2.\lVert y'-y_G\rVert_2\le Cm^3\rho,\qquad \left|\sqrt{m}|y'_j|-\sqrt{m}|(y_G)_j|\right|\le Cm^{7/2}\rho,\qquad j=1,2.

The Gaussian kernel line is Haar. Each scaled coordinate magnitude of a Haar vector in Rm\mathbb{R}^m has a density bounded uniformly for sufficiently large mm. Consequently,

PG(G∈∂ρΩ)≤Cm−3+Ce−c′m+Cm7/2ρ=o(1).\mathbb{P}_{G}(G\in\partial_\rho\Omega)\le Cm^{-3}+Ce^{-c'm}+Cm^{7/2}\rho=o(1).

Here m7/2ρ≤n−43c/200m^{7/2}\rho\le n^{-43c/200}. Combining this estimate with (A.5), and then adding the probability of omitted-column irregularity, proves

(m∣y1∣,m∣y2∣)⟹(∣g1∣,∣g2∣).(76)(\sqrt{m}|y_1|,\sqrt{m}|y_2|)\Longrightarrow(|g_1|,|g_2|). \tag*{(76)}

One may define yy arbitrarily on failure of full row rank. When A\mathcal{A} and the omitted block both have full rank, the rank decomposition under projection implies rank⁡Z=m−1\operatorname{rank} Z=m-1. Their exceptional probabilities are exponentially small.

Return to the original coordinates. On this full-rank event, vI≠0v_I\ne0: otherwise the full-column-rank omitted block would annihilate the nonzero vector vIcv_{I^c}. Projecting Av=0\mathcal{A}v=0 shows that the kernel line of ZZ is spanned by vIv_I, so y=vI/∥vI∥2y=v_I/\lVert v_I\rVert_2 up to sign. Therefore

(n∣v1∣,n∣v2∣)=Sn(m∣y1∣,m∣y2∣).(\sqrt{n}|v_1|,\sqrt{n}|v_2|)=\sqrt{S_n}(\sqrt{m}|y_1|,\sqrt{m}|y_2|).

Equations (A.3) and (A.7), together with Slutsky’s theorem, give the required limit. Independence between SnS_n and yy is unnecessary. Finally, the Haar-line extension on rank deficiency changes none of these limits.

References

References

  1. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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

Paper details

vlnc:W_dmt5nx2u
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents