betaFeedback

WorksMathematics

Spectral outliers of self-coupled covariance matrices with exact zero weights

Abstract

We study the right spectral outliers of Md=n−1AdDAdTM_d=n^{-1}A_dDA_d^{\mathsf T}, where the columns of AdA_d are drawn from a Gaussian mixture with a fixed-dimensional signal subspace and each diagonal weight Dℓℓ=hbℓ(gℓ)D_{\ell\ell}=h_{b_\ell}(g_\ell) is a function of the class label and of the low-dimensional Gaussian coordinates of the same sample. The effective outlier theory of Ben Arous, Gheissari, Huang and Jagannath for such self-coupled matrices assumes that the weights do not accumulate at zero, which excludes hard gates such as the derivative of a rectified linear unit. Assuming hb(g)∈{0}∪[τ,M]h_b(g)\in\{0\}\cup[\tau,M], we prove that exact zero weights do not change the positive right-outlier equation. The proof deletes inactive columns before any linearization, keeps their random normalization, and applies the invertible-weight theory to the exact conditional mark law of the active samples; we show, by conditioning on the marks, that this theory extends to non-Gaussian fixed-dimensional marks. The limiting spectrum acquires the atom (1−ϕπ)+(1-\phi\pi)_+ at zero, while positive outliers are the roots of a q×qq\times q determinant. We obtain multiplicity-preserving eigenvalue counts, absence of spurious right outliers, a projected-eigenvector normalization, an invariant-subspace formula at repeated roots, and the exact finite-dimensional kernel.

Introduction

Random covariance spectra separate macroscopic bulk behaviour from finitely many signal-bearing outliers. In classical spiked models this separation is controlled by deterministic transforms of the bulk law and produces the Baik–Ben Arous–Péché transition [8, 2, 4]. In self-coupled models, the diagonal weight of a sample depends on the same low-dimensional projection that enters the sample itself. Such matrices describe spectral estimators, information matrices, and blocks of neural-network curvature [3, 7].

An effective spectral theory for Gaussian-mixture self-coupling was developed by Ben Arous, Gheissari, Huang and Jagannath [3]. Its outlier proof linearizes the weighted covariance matrix using the inverse of the diagonal weights, and its assumptions forbid weights accumulating at zero. Consequently a hard gate, for example the derivative of a rectified linear unit (ReLU) [10], which vanishes on a set of positive probability, is excluded. We show that this singularity is algebraic rather than spectral: inactive columns can be removed exactly, without changing the effective equation for positive outliers.

Exact zero weights are already covered in a related centred setting. For isotropic standard Gaussian covariates, Kovačević, Zhang and Mondelli [7] treat spectral estimators n−1∑iT(yi)aiaiTn^{-1}\sum_i T(y_i)a_i a_i^{\mathsf{T}} for multi-index models with a bounded preprocessing function TT subject only to P{T(y)=0}<1\mathbb{P}\{T(y)=0\}<1, so that the weights may vanish with positive probability; they locate the top eigenvalues and determine the eigenvector overlaps with the signal subspace. The present note treats the noncentred Gaussian-mixture model of [3], and gives multiplicity-preserving counts on arbitrary compact intervals to the right of the bulk. Recent local-law and block-Wishart results [5, 9] provide broader context.

The argument has two parts. First, an exact finite-dimensional identity (Section 3) writes MdM_d as a random scalar multiple of an invertible-weight matrix built from the active columns only; these columns are i.i.d. from a truncated, generally non-Gaussian, mark law. Second, we show (Proposition 4.1) that the invertible-weight theory of [3] holds for arbitrary fixed-dimensional marks that are independent of the orthogonal Gaussian block: conditionally on the marks, every estimate used in the source proof concerns either the Gaussian block with a fixed weight matrix, or laws of large numbers for mark functionals.

Scope. All statements are asymptotic in the proportional regime n/d→ϕn/d \to\phi with fixed summary dimension qq and fixed number of classes kk. They require nonnegative weights and a fixed gap τ>0\tau> 0 between zero and the nonzero weights. Signed weights, positive weights accumulating continuously at zero, growing summary dimension, and uniformity along a data-dependent training trajectory are not covered. The numerical computations in § 12 are illustrations and are not used in any proof. Whether the conclusions persist when the gap τ=τd\tau= \tau_d tends to zero with dd, and at what quantitative rate the empirical outliers approach the effective roots, are open questions.

Model and main theorem

Fix k,C,q∈Nk,C,q \in\mathbb{N} with q=C+kq=C+k, λ>0\lambda> 0, ϕ>0\phi> 0, probabilities pb>0p_b > 0 summing to one, and a positive-semidefinite matrix G∈Rq×qG \in\mathbb{R}^{q \times q}. For each large dd, let

Bd=[x1(d),…,xC(d),μ1(d),…,μk(d)](1)B_d = [x_1^{(d)},\ldots,x_C^{(d)},\mu_1^{(d)},\ldots,\mu_k^{(d)}] \tag*{(1)}

satisfy BdTBd=GB_d^{\mathsf{T}}B_d=G. Choose Ld∈Rd×qL_d \in\mathbb{R}^{d \times q} with orthonormal columns and

LdTBd=G.(2)L_d^{\mathsf{T}}B_d=\sqrt{G}. \tag*{(2)}

No inverse of GG or G\sqrt{G} is used. If GG is rank deficient, unused columns of LdL_d may be chosen arbitrarily outside span⁡Bd\operatorname{span} B_d. Let Ld⊥∈Rd×(d−q)L_d^\perp\in\mathbb{R}^{d \times(d-q)} complete LdL_d to an orthogonal matrix.

For class bb, let mb=LdTμb(d)m_b=L_d^{\mathsf{T}}\mu_b^{(d)}. Independently over ℓ\ell sample

bℓ∼Cat⁡(p1,…,pk),gℓ∼N(0,Iq),wℓ∼N(0,Id−q),(3)b_\ell\sim\operatorname{Cat}(p_1,\ldots,p_k),\qquad g_\ell\sim\mathcal{N}(0,I_q),\qquad w_\ell\sim\mathcal{N}(0,I_{d-q}), \tag*{(3)}

and define

Yℓ=Ld(mbℓ+λ−1/2gℓ)+λ−1/2Ld⊥wℓ.(4)Y_\ell=L_d\left(m_{b_\ell}+\lambda^{-1/2}g_\ell\right)+\lambda^{-1/2}L_d^\perp w_\ell. \tag*{(4)}

Let Ad=[Y1 ⋯ Yn]A_d=[Y_1\ \cdots\ Y_n], with n/d→ϕn/d \to\phi, and let

Dℓℓ=hbℓ(gℓ),Md=1nAdDAdT,(5)D_{\ell\ell}=h_{b_\ell}(g_\ell),\qquad M_d=\frac{1}{n}A_dD A_d^{\mathsf{T}}, \tag*{(5)}

where hb:Rq→[0,M]h_b:\mathbb{R}^q \to[0,M] is measurable and

hb(g)∈{0}∪[τ,M]a.s.(6)h_b(g)\in\{0\}\cup[\tau,M]\quad\text{a.s.} \tag*{(6)}

for some τ>0\tau>0. Put

πb=P{hb(g)>0},π=∑b=1kpbπb.(7)\pi_b=\mathbb{P}\{h_b(g)>0\},\qquad\pi=\sum_{b=1}^{k}p_b\pi_b. \tag*{(7)}

Let SGS_G be the physical Stieltjes branch

1+zSG(z)=ϕ∑b=1kpbE[SG(z)hb(g)λϕ+SG(z)hb(g)],(8)1+zS_G(z)=\phi\sum_{b=1}^{k}p_b\mathbb{E}\left[\frac{S_G(z)h_b(g)}{\lambda\phi+S_G(z)h_b(g)}\right], \tag*{(8)}

with Im⁡SG(z)>0\operatorname{Im}S_G(z)>0 for Im⁡z>0\operatorname{Im}z>0 and zSG(z)→−1zS_G(z)\to-1. Let νG\nu_G be its probability measure and

λ+(G)=sup⁡supp⁡νG.(9)\lambda_+(G)=\sup\operatorname{supp}\nu_G. \tag*{(9)}

For

vb(g)=mb+λ−1/2g,(10)v_b(g)=m_b+\lambda^{-1/2}g, \tag*{(10)}

define, outside the support,

FG(z)=λϕ∑b=1kpbE[hb(g)λϕ+SG(z)hb(g)vb(g)vb(g)T],(11)F_G(z)=\lambda\phi\sum_{b=1}^{k}p_b\mathbb{E}\left[\frac{h_b(g)}{\lambda\phi+S_G(z)h_b(g)}v_b(g)v_b(g)^{\mathsf{T}}\right], \tag*{(11)}

and BG(z)=zIq−FG(z)B_G(z)=zI_q-F_G(z).

Theorem 2.1 (Zero-atom outlier theorem). If π=0\pi=0, then Md=0M_d=0 almost surely and all assertions reduce to νG=δ0\nu_G=\delta_0, SG(z)=−1/zS_G(z)=-1/z, and FG=0F_G=0. Assume π>0\pi>0.

(1) The empirical spectral distribution of MdM_d converges weakly in probability to νG\nu_G, and

νG({0})=(1−ϕπ)+.(12)\nu_G(\{0\})=(1-\phi\pi)_+. \tag*{(12)}

(2) For every compact K⊂(λ+(G),∞)K\subset(\lambda_+(G),\infty),

ess inf⁡z∈K, b∈[k]g:hb(g)>0∣λϕ+SG(z)hb(g)∣>0.(13)\underset{\substack{z\in K,\ b\in[k]\\ g:h_b(g)>0}}{\operatorname{ess\,inf}}\left|\lambda\phi+S_G(z)h_b(g)\right|>0. \tag*{(13)}

Hence FGF_G and BGB_G are real analytic to the right of the bulk.

(3) Let I=[a,b]⊂(λ+(G),∞)I=[a,b]\subset(\lambda_+(G),\infty) and assume det⁡BG(a)det⁡BG(b)≠0\det B_G(a)\det B_G(b)\ne0. Then, with probability tending to one,

#{σ(Md)∩I}=#{z∈I:det⁡BG(z)=0},(14)\#\{\sigma(M_d)\cap I\}=\#\{z\in I:\det B_G(z)=0\}, \tag*{(14)}

with algebraic and analytic multiplicity.

(4) For each root z∗>λ+(G)z_*> \lambda_+(G)*,

ord⁡z∗det⁡BG(z)=dim⁡Ker⁡BG(z∗).(15)\operatorname{ord}_{z_*}\det B_G(z)=\dim\operatorname{Ker}B_G(z_*). \tag*{(15)}

The sum of right-root multiplicities is at most qq.

(5) Every isolated root of multiplicity mm attracts exactly mm empirical eigenvalues. Conversely, for every ϵ>0\epsilon>0, all empirical eigenvalues above λ+(G)+ϵ\lambda_+(G)+\epsilon lie within oP(1)o_{\mathbb{P}}(1) of a bounded effective root.

(6) If z∗z_* is simple, zd→z∗z_d\to z_* is its empirical eigenvalue, udu_d is a unit eigenvector, and rd=LdTudr_d=L_d^\mathsf{T}u_d, then

∥BG(zd)rd∥=oP(1),(16)\lVert B_G(z_d)r_d\rVert=o_{\mathbb{P}}(1), \tag*{(16)}
rdTBG′(zd)rd−1=oP(1).(17)r_d^\mathsf{T}B_G'(z_d)r_d-1=o_{\mathbb{P}}(1). \tag*{(17)}

If e∗e_* is a unit vector spanning Ker⁡BG(z∗)\operatorname{Ker}B_G(z_*), then after sign choice,

rd→Pe∗e∗TBG′(z∗)e∗.(18)r_d\xrightarrow{\mathbb{P}}\frac{e_*}{\sqrt{e_*^\mathsf{T}B_G'(z_*)e_*}}. \tag*{(18)}

(7) If z∗z_* has multiplicity m>1m>1, choose an isolating interval and let Πd\Pi_d be the corresponding empirical spectral projection. Let

K∗=Ker⁡BG(z∗),P∗=Proj⁡K∗,J∗=P∗BG′(z∗)P∗∣K∗.(19)K_*=\operatorname{Ker}B_G(z_*),\qquad P_*=\operatorname{Proj}_{K_*},\qquad J_*=P_*B_G'(z_*)P_*\big|_{K_*}. \tag*{(19)}

Then J∗≻0J_*\succ0, rank Πd=m\Pi_d=m with probability tending to one, and

LdTΠdLd→PP∗J∗−1P∗.(20)L_d^\mathsf{T}\Pi_dL_d\xrightarrow{\mathbb{P}}P_*J_*^{-1}P_*. \tag*{(20)}

(8) For every finite dd, almost surely,

rank⁡Md=min⁡(d,Nactive),Nactive=∑ℓ=1n1{Dℓℓ>0}.(21)\operatorname{rank}M_d=\min(d,N_{\mathrm{active}}),\qquad N_{\mathrm{active}}=\sum_{\ell=1}^{n}\mathbf{1}_{\{D_{\ell\ell}>0\}}. \tag*{(21)}

Consequently,

1ddim⁡Ker⁡Md⟶(1−ϕπ)+almost surely.(22)\frac{1}{d}\dim\operatorname{Ker}M_d\longrightarrow(1-\phi\pi)_+\quad\text{almost surely}. \tag*{(22)}

The remainder of the note proves Theorem 2.1. § 3–§ 4 contain the reduction and the extension of the invertible-weight theory; § 5–§ 10 derive the individual assertions.

Exact active-column reduction

Define

Tℓ=hbℓ(gℓ),Iℓ=1{Tℓ>0},Nd=∑ℓ=1nIℓ.(23)T_\ell=h_{b_\ell}(g_\ell),\qquad I_\ell=\mathbf{1}_{\{T_\ell>0\}},\qquad N_d=\sum_{\ell=1}^{n}I_\ell. \tag*{(23)}

After a column permutation,

Ad=[A+,d A0,d],D=[D+,d000],(24)A_d=[A_{+,d}\ A_{0,d}],\qquad D=\begin{bmatrix}D_{+,d}&0\\0&0\end{bmatrix}, \tag*{(24)}

where D+,dD_{+,d} is Nd×NdN_d \times N_d and

τI⪯D+,d⪯MI.(25)\tau I \preceq D_{+,d} \preceq MI. \tag*{(25)}

Therefore

Md=1nA+,dD+,dA+,dT=NdnM^d,M^d=1NdA+,dD+,dA+,dT.(26)M_d=\frac{1}{n}A_{+,d}D_{+,d}A_{+,d}^{\mathsf T}=\frac{N_d}{n}\widehat{M}_d,\qquad\widehat{M}_d=\frac{1}{N_d}A_{+,d}D_{+,d}A_{+,d}^{\mathsf T}. \tag*{(26)}

This is an exact finite-dimensional identity. The inactive columns are absent before any determinant identity, Schur complement, or resolvent is introduced.

Lemma 3.1 (Conditional active law). Conditioned on activity, the low-dimensional mark has law

P+(b,dg)=pb1{hb(g)>0}γq(dg)π,(27)\mathbb{P}_{+}(b,dg)=\frac{p_b\mathbf{1}_{\{h_b(g)>0\}}\gamma_q(dg)}{\pi}, \tag*{(27)}

and ww remains an independent standard Gaussian. Conditioned on Nd=mN_d=m, the unordered active columns are distributed as mm independent draws from this active law.

Proof. The activity event is measurable with respect to (b,g)(b,g) and independent of ww. Thus

P{b∈db,g∈dg,w∈dw∣I=1}=pb1{hb(g)>0}γq(dg)πγd−q(dw).(28)\mathbb{P}\{b\in db,g\in dg,w\in dw\mid I=1\}=\frac{p_b\mathbf{1}_{\{h_b(g)>0\}}\gamma_q(dg)}{\pi}\gamma_{d-q}(dw). \tag*{(28)}

Independence across sample indices is preserved under coordinatewise conditioning. Given the number of successes, exchangeability makes their positions uniform, and the unordered successful marks are independent with law (27). □

Remark 3.2. The active gg is generally truncated or reweighted. It must not be replaced by an unconditioned Gaussian vector: replacing (27) by the unconditioned law would alter the low-rank term. Thus the active columns are not, in general, an ordinary Gaussian mixture.

Because IℓI_\ell are independent Bernoulli variables of mean π\pi,

Ndn→π,Ndd→ϕπ(29)\frac{N_d}{n}\to\pi,\qquad\frac{N_d}{d}\to\phi\pi \tag*{(29)}

almost surely, for example by Hoeffding’s inequality and Borel–Cantelli.

Invertible covariance with fixed-dimensional marks

The invertible-weight theory of [3] treats a weighted covariance matrix whose weights do not accumulate at zero and whose low-dimensional coordinates are Gaussian. After the reduction of § 3 the weights are bounded below, but the low-dimensional mark of an active column has the truncated law (27). This section shows that Gaussianity of the mark is not used.

Statement

Fix qq, λ>0\lambda>0 and 0<τ≤M0<\tau\le M. Let μ\mu be a probability law on [τ,M]×Rq[\tau,M]\times\mathbb{R}^{q} with

∫∥v∥r μ(dt,dv)<∞for every r<∞,(30)\int\lVert v\rVert^r\,\mu(dt,dv)<\infty\qquad\text{for every }r<\infty, \tag*{(30)}

and let μT\mu_T denote its first marginal. For each dd let md∈Nm_d\in\mathbb{N} with md/d→c∈(0,∞)m_d/d\to c\in(0,\infty), let L=Ld∈Rd×qL=L_d\in\mathbb{R}^{d\times q} have orthonormal columns, and let L⊥L^\perp complete it to an orthogonal matrix. Let (Ti,Vi)i≤md(T_i,V_i)_{i\le m_d} be i.i.d. with law μ\mu, and let Wi∼N(0,Id−q)W_i\sim\mathcal{N}(0,I_{d-q}) be i.i.d. and independent of the marks. Define

Yi=LVi+λ−1/2L⊥Wi,M^d=1md∑i=1mdTiYiYiT.(31)Y_i=LV_i+\lambda^{-1/2}L^\perp W_i,\qquad\widehat{M}_d=\frac{1}{m_d}\sum_{i=1}^{m_d}T_iY_iY_i^{\mathsf T}. \tag*{(31)}

Let scs_c be the Stieltjes transform of the probability measure νc\nu_c determined by

1+xsc(x)=cE[sc(x)Tλc+sc(x)T],(32)1+xs_c(x)=c\mathbb{E}\left[\frac{s_c(x)T}{\lambda c+s_c(x)T}\right], \tag*{(32)}

with Im⁡sc>0\operatorname{Im}s_c>0 on C+\mathbb{C}^{+}, let λ+,c=sup⁡supp⁡νc\lambda_{+,c}=\sup\operatorname{supp}\nu_c, and set

Fc(x)=λcE[Tλc+sc(x)TVVT],Bc(x)=xIq−Fc(x).(33)F_c(x)=\lambda c\mathbb{E}\left[\frac{T}{\lambda c+s_c(x)T}VV^{\mathsf T}\right],\qquad B_c(x)=xI_q-F_c(x). \tag*{(33)}

Expectations are over (T,V)∼μ(T,V)\sim\mu.

Almost-sure statements refer to one probability space carrying the variables for all dd, on which the family of all marks (over all dd) is independent of the family of all Gaussian vectors (over all dd). The marks may be coupled across dd arbitrarily; the Gaussian vectors are coupled across dd in any way under which the almost-sure statements of [3] for a Gaussian block independent of a deterministic weight matrix hold (the source does not fix a coupling). Every later use of the proposition (§ 8 and § 10) is through probabilities at fixed dd, which depend only on the law at that dd and not on these couplings.

Proposition 4.1 (Invertible covariance with fixed-dimensional marks). In this setting the following hold.

(i) The empirical spectral distribution of M^d\widehat{M}_d converges weakly to νc\nu_c almost surely.

(ii) For every compact K⊂(λ+,c,∞)K\subset(\lambda_{+,c},\infty), inf⁡{∣λc+sc(x)t∣:x∈K, t∈supp⁡μT}>0\inf\{|\lambda_c+s_c(x)t|:x\in K,\ t\in\operatorname{supp}\mu_T\}>0. Hence FcF_c and BcB_c are real analytic on (λ+,c,∞)(\lambda_{+,c},\infty).

(iii) Let [a,b]⊂(λ+,c,∞)[a,b]\subset(\lambda_{+,c},\infty) with det⁡Bc(a)det⁡Bc(b)≠0\det B_c(a)\det B_c(b)\ne0. Almost surely, for all large dd, the number of eigenvalues of M^d\widehat{M}_d in [a,b][a,b], counted with multiplicity, equals the number of zeros of det⁡Bc\det B_c in [a,b][a,b], counted with order.

(iv) For every [a,b]⊂(λ+,c,∞)[a,b]\subset(\lambda_{+,c},\infty) there is η>0\eta>0 such that, with Ω={x∈C:Re⁡x∈[a,b], ∣Im⁡x∣≤η}\Omega=\{x\in\mathbb{C}:\operatorname{Re}x\in[a,b],\ |\operatorname{Im}x|\le\eta\}, there are random q×qq\times q matrix functions BdB_d, holomorphic on a neighbourhood of Ω\Omega for all large dd, with

LT(M^d−x)−1L=−Bd(x)−1(x∈Ω∖σ(M^d)),(34)L^{\mathsf T}(\widehat{M}_d-x)^{-1}L=-B_d(x)^{-1}\qquad(x\in\Omega\setminus\sigma(\widehat{M}_d)), \tag*{(34)}

and, almost surely, sup⁡Ω∥Bd−Bc∥→0\sup_{\Omega}\|B_d-B_c\|\to0 and sup⁡Ω∥Bd′−Bc′∥→0\sup_{\Omega}\|B'_d-B'_c\|\to0.

(v) *If xd∈σ(M^d)∩[a,b]x_d\in\sigma(\widehat{M}_d)\cap[a,b], xd→x∗x_d\to x_*, udu_d is a unit eigenvector and rd=LTudr_d=L^{\mathsf T}u_d, then almost surely ∥Bc(xd)rd∥→0\|B_c(x_d)r_d\|\to0 and rdTBc′(xd)rd→1r_d^{\mathsf T}B'_c(x_d)r_d\to1.

Two laws of large numbers

The source proof invokes the strong law of large numbers for triangular arrays. We record the version used here; it requires no coupling between different values of dd.

Lemma 4.2. Let ξ\xi be a real random variable with E∣ξ∣6<∞\mathbb{E}|\xi|^6<\infty. For each dd let ξd,1,…,ξd,md\xi_{d,1},\ldots,\xi_{d,m_d} be i.i.d. copies of ξ\xi, with md/d→c>0m_d/d\to c>0. Then md−1∑i≤mdξd,i→Eξm_d^{-1}\sum_{i\le m_d}\xi_{d,i}\to\mathbb{E}\xi almost surely.

Proof. Put ζ=ξ−Eξ\zeta=\xi-\mathbb{E}\xi and Σm=∑i≤mζd,i\Sigma_m=\sum_{i\le m}\zeta_{d,i}. In the expansion of EΣm6\mathbb{E}\Sigma_m^6 over index tuples (i1,…,i6)(i_1,\ldots,i_6), a tuple containing an index exactly once contributes zero. The remaining tuples take at most three distinct values, so there are at most C0m3C_0m^3 of them for an absolute constant C0C_0, and each contributes at most E∣ζ∣6\mathbb{E}|\zeta|^6 by Hölder’s inequality. Markov’s inequality gives P{∣Σmd∣>ϵmd}≤C0E∣ζ∣6ϵ−6md−3\mathbb{P}\{|\Sigma_{m_d}|>\epsilon m_d\}\le C_0\mathbb{E}|\zeta|^6\epsilon^{-6}m_d^{-3}, which is summable in dd because md≥cd/2m_d\ge cd/2 eventually. The Borel–Cantelli lemma completes the proof. □\square

Lemma 4.3. Let Td,1,…,Td,mdT_{d,1},\ldots,T_{d,m_d} be i.i.d. with law μT\mu_T supported in [τ,M][\tau,M], and let μ^d=md−1∑iδTd,i\widehat{\mu}_d=m_d^{-1}\sum_i\delta_{T_{d,i}}. Almost surely, μ^d→μT\widehat{\mu}_d\to\mu_T weakly and the Hausdorff distance between supp⁡μ^d\operatorname{supp}\widehat{\mu}_d and supp⁡μT\operatorname{supp}\mu_T tends to zero.

Proof. Lemma 4.2 applied to the countably many bounded variables TjT^j, j∈Nj\in\mathbb{N}, gives almost sure convergence of all moments; on [τ,M][\tau,M] this implies weak convergence by the Weierstrass approximation theorem. Almost surely every Td,iT_{d,i} lies in supp⁡μT\operatorname{supp}\mu_T, so supp⁡μ^d⊂supp⁡μT\operatorname{supp}\widehat{\mu}_d\subset\operatorname{supp}\mu_T. Conversely, fix ϵ>0\epsilon>0 and cover the compact set supp⁡μT\operatorname{supp}\mu_T by finitely many open intervals J1,…,JrJ_1,\ldots,J_r of length ϵ\epsilon centred at support points. Each has pj=μT(Jj)>0p_j=\mu_T(J_j)>0, and the probability that some JjJ_j contains no sample is at most r(1−min⁡jpj)mdr(1-\min_jp_j)^{m_d}, which is summable in dd. By Borel–Cantelli, eventually every point of supp⁡μT\operatorname{supp}\mu_T is within ϵ\epsilon of a sample. Take ϵ=1/k\epsilon=1/k, k∈Nk\in\mathbb{N}. □\square

The finite-d transform for a convergent aspect ratio

The source’s finite-dd transform S(d)S^{(d)} solves its equation (2.11), which carries the ratio ϕd=n/(d−q)\phi_d=n/(d-q); this ratio is not constant even when n/d=ϕn/d=\phi. Here ϕd=md/(d−q)→c\phi_d=m_d/(d-q)\to c. The following lemma gives the convergence on C+\mathbb{C}^+ in this generality.

Lemma 4.4. Let ρd,ρ\rho_d,\rho be probability measures on [τ,M][\tau,M] with ρd→ρ\rho_d\to\rho weakly, and let ϕd,c>0\phi_d,c>0 with ϕd→c\phi_d\to c. For z∈C+z\in\mathbb{C}^+ and a probability measure ρ\rho on [τ,M][\tau,M], γ>0\gamma>0, consider

1S=−z+ϕ∫t1+St ρ(dt),Im⁡S>0.(35)\frac{1}{S}=-z+\phi\int\frac{t}{1+St}\,\rho(dt),\qquad\operatorname{Im}S>0. \tag*{(35)}

(a) (35) has at most one solution, and every solution satisfies ∣S∣≤1/Im⁡z|S| \le1/\operatorname{Im} z and φ∫t2∣1+St∣−2ϱ(dt)<∣S∣−2\varphi\int t^2|1+St|^{-2}\varrho(dt) < |S|^{-2}.

(b) If Sd(z)S_d(z) solves (35) with (ϱ,φ)=(ρd,ϕd)(\varrho,\varphi)=(\rho_d,\phi_d) and S(z)S(z) solves it with (ρ,c)(\rho,c), then Sd(z)→S(z)S_d(z) \to S(z) for every z∈C+z \in\mathbb{C}^{+}. If moreover the SdS_d are holomorphic on C+\mathbb{C}^{+}, the convergence is uniform on compact subsets of C+\mathbb{C}^{+}.

Proof. For Im⁡S>0\operatorname{Im} S>0 and t>0t>0, Im⁡(1+St)=tIm⁡S>0\operatorname{Im}(1+St)=t\operatorname{Im} S>0, so the integrand is defined. Since Im⁡(1/S)=−Im⁡S/∣S∣2\operatorname{Im}(1/S)=-\operatorname{Im} S/|S|^2 and Im⁡(t/(1+St))=−t2Im⁡S/∣1+St∣2\operatorname{Im}(t/(1+St))=-t^2\operatorname{Im} S/|1+St|^2, the imaginary part of (35) reads

Im⁡S(1∣S∣2−φ∫t2∣1+St∣2ϱ(dt))=Im⁡z>0.(36)\operatorname{Im} S\left(\frac{1}{|S|^2}-\varphi\int\frac{t^2}{|1+St|^2}\varrho(dt)\right)=\operatorname{Im} z>0. \tag*{(36)}

This gives the integral bound in (a), and also Im⁡S≥∣S∣2Im⁡z\operatorname{Im} S \ge|S|^2\operatorname{Im} z, hence ∣S∣≤1/Im⁡z|S|\le1/\operatorname{Im} z. If S1,S2S_1,S_2 are solutions, subtracting the two equations gives

(S1−S2)(1S1S2−φ∫t2(1+S1t)(1+S2t)ϱ(dt))=0.(S_1-S_2)\left(\frac{1}{S_1S_2}-\varphi\int\frac{t^2}{(1+S_1t)(1+S_2t)}\varrho(dt)\right)=0.

By the Cauchy–Schwarz inequality and the integral bound, the modulus of the integral term is strictly less than 1/(∣S1∣∣S2∣)1/(|S_1||S_2|), so the bracket is nonzero and S1=S2S_1=S_2.

For (b), fix zz and put Φ=sup⁡dϕd<∞\Phi=\sup_d\phi_d<\infty. If ∣Sd∣≤1/(2M)|S_d|\le1/(2M), then ∣1+Sdt∣≥1/2|1+S_dt|\ge1/2 on [τ,M][\tau,M] and (35) gives 1/∣Sd∣≤∣z∣+2ΦM1/|S_d|\le|z|+2\Phi M. Hence ∣Sd∣≥s0=min⁡{1/(2M),1/(∣z∣+2ΦM)}|S_d|\ge s_0=\min\{1/(2M),1/(|z|+2\Phi M)\}, and by (36), Im⁡Sd≥s02Im⁡z\operatorname{Im} S_d\ge s_0^2\operatorname{Im} z. Consequently ∣1+Sdt∣≥tIm⁡Sd≥κ=τs02Im⁡z|1+S_dt|\ge t\operatorname{Im} S_d\ge\kappa=\tau s_0^2\operatorname{Im} z for t∈[τ,M]t\in[\tau,M]. The sequence SdS_d is bounded by (a); let S∞S_\infty be a limit point along a subsequence. Then Im⁡S∞≥s02Im⁡z>0\operatorname{Im} S_\infty\ge s_0^2\operatorname{Im} z>0, so ∣1+S∞t∣≥κ|1+S_\infty t|\ge\kappa on [τ,M][\tau,M], and

sup⁡t∈[τ,M]∣t1+Sdt−t1+S∞t∣≤M2∣Sd−S∞∣κ2→0\sup_{t\in[\tau,M]}\left|\frac{t}{1+S_dt}-\frac{t}{1+S_\infty t}\right|\le\frac{M^2|S_d-S_\infty|}{\kappa^2}\to0

along the subsequence. Since t↦t/(1+S∞t)t\mapsto t/(1+S_\infty t) is bounded and continuous on [τ,M][\tau,M] and ρd→ρ\rho_d\to\rho weakly,

∫t(1+Sdt)−1ρd(dt)→∫t(1+S∞t)−1ρ(dt).\int t(1+S_dt)^{-1}\rho_d(dt)\to\int t(1+S_\infty t)^{-1}\rho(dt).

With ϕd→c\phi_d\to c, S∞S_\infty solves (35) for (ρ,c)(\rho,c) and has positive imaginary part, so S∞=S(z)S_\infty=S(z) by (a). Every subsequence has a further subsequence converging to S(z)S(z), so Sd(z)→S(z)S_d(z)\to S(z). Under holomorphy, the bound ∣Sd∣≤1/Im⁡z|S_d|\le1/\operatorname{Im} z makes (Sd)(S_d) locally bounded on C+\mathbb{C}^{+}, and Vitali’s theorem upgrades pointwise to locally uniform convergence.

The next lemma controls the right edge of the finite-dd measures; it is used only in the paragraph Avoiding the modification for (ii)–(v) below.

Lemma 4.5. Let ρd\rho_d be finitely supported probability measures on [τ/2,s][\tau/2,s], where s=max⁡supp⁡μTs=\max\operatorname{supp}\mu_T, with ρd→μT\rho_d\to\mu_T weakly, and let ϕd→c>0\phi_d\to c>0. Let νd\nu_d and ν\nu be the probability measures whose Stieltjes transforms solve (35) with (ρd,ϕd)(\rho_d,\phi_d) and with (μT,c)(\mu_T,c), respectively. Then lim sup⁡dsup⁡supp⁡νd≤sup⁡supp⁡ν\limsup_d\sup\operatorname{supp}\nu_d\le\sup\operatorname{supp}\nu.

Proof. For a measure ϱ\varrho and ϕ>0\phi>0 write fϱ,ϕ(y)=−1/y+ϕ∫t(1+yt)−1ϱ(dt)f_{\varrho,\phi}(y)=-1/y+\phi\int t(1+yt)^{-1}\varrho(dt). Let x>sup⁡supp⁡νx>\sup\operatorname{supp}\nu and y=S(x)=∫ν(du)/(u−x)<0y=S(x)=\int\nu(du)/(u-x)<0. By the Silverstein–Choi characterization used in the paragraph Denominators below, 1+S(x′)t≠01+S(x')t\ne0 for x′≥xx'\ge x and t∈supp⁡μTt\in\operatorname{supp}\mu_T; since SS is continuous on [x,∞)[x,\infty) and S(x′)→0S(x')\to0 as x′→∞x'\to\infty, in fact 1+yt>01+yt>0 on supp⁡μT\operatorname{supp}\mu_T, so y∈(−1/s,0)y\in(-1/s,0). Letting z→xz\to x in (35) gives x=fμT,c(y)x=f_{\mu_T,c}(y). Now let sd=max⁡supp⁡ρd≤ss_d=\max\operatorname{supp}\rho_d\le s. By [6 Lemmas 2.4–2.6], applied with τ=ρd\tau=\rho_d, ϕd\phi_d, sup⁡supp⁡νd≤max⁡{0,fρd,ϕd(x1)}\sup\operatorname{supp}\nu_d\le\max\{0,f_{\rho_d,\phi_d}(x_1)\}, where x1x_1 is the unique critical point of fρd,ϕdf_{\rho_d,\phi_d} in (−1/sd,0)(-1/s_d,0). This function tends to +∞+\infty at both ends of that interval, because ρd\rho_d has an atom at sds_d, so fρd,ϕd(x1)f_{\rho_d,\phi_d}(x_1) is its minimum there. Since y∈(−1/s,0)⊂(−1/sd,0)y\in(-1/s,0)\subset(-1/s_d,0),

sup⁡supp⁡νd≤max⁡{0,fρd,ϕd(y)}.\sup\operatorname{supp}\nu_d\le\max\{0,f_{\rho_d,\phi_d}(y)\}.

On [τ/2,s][\tau/2,s] the function t↦t/(1+yt)t\mapsto t/(1+yt) is bounded and continuous, because 1+yt≥1+ys>01+yt\ge1+ys>0. Weak convergence and ϕd→c\phi_d\to c give fρd,ϕd(y)→fμT,c(y)=x>0f_{\rho_d,\phi_d}(y)\to f_{\mu_T,c}(y)=x>0. Hence lim sup⁡dsup⁡supp⁡νd≤x\limsup_d\sup\operatorname{supp}\nu_d\le x, and x↓sup⁡supp⁡νx\downarrow\sup\operatorname{supp}\nu completes the proof.

Proof of Proposition 4.1

All references in this proof are to the numbered statements of [3], version 3, Sections 1.4 and 2.

Normalization. The source works with Xd=(λ/d)∑iTiYiYiT=λcdM^dX_d=(\lambda/d)\sum_i T_iY_iY_i^{\mathsf T}=\lambda c_d\widehat{M}_d, where cd=md/dc_d=m_d/d. Write

λdYi=Lri+L⊥Wid,ri=λdVi.(37)\sqrt{\frac{\lambda}{d}}Y_i=Lr_i+L^\perp\frac{W_i}{\sqrt{d}},\qquad r_i=\sqrt{\frac{\lambda}{d}}V_i. \tag*{(37)}

Let R∈Rmd×qR\in\mathbb{R}^{m_d\times q} have rows riTr_i^{\mathsf T}, let W∈R(d−q)×mdW\in\mathbb{R}^{(d-q)\times m_d} have columns Wi/dW_i/\sqrt{d}, and let D=diag⁡(T1,…,Tmd)D=\operatorname{diag}(T_1,\ldots,T_{m_d}). This is the decomposition (2.34)–(2.36) of the source, with the Gaussian vector g+λLTμag+\sqrt{\lambda}L^{\mathsf T}\mu_a there replaced by λVi\sqrt{\lambda}V_i.

The conditioning. Let G\mathcal{G} be the σ\sigma-field generated by the marks (Ti(d),Vi(d))(T_i^{(d)},V_i^{(d)}) for all dd and i≤mdi\le m_d. For each dd, D=D(d)D=D^{(d)} and R=R(d)R=R^{(d)} are G\mathcal{G}-measurable, and, by the independence of the two families, conditionally on G\mathcal{G} the sequence (W(d))d(W^{(d)})_d has its unconditional joint law; in particular each W(d)W^{(d)} has i.i.d. N(0,1/d)N(0,1/d) entries. In the source, Assumption 1 (weights are functions of the label and of the Gaussian summary coordinates) enters the proof of Theorem 2.1 only through this structure and through the computation of a mean: it is used in Section 2.6 to conclude that WW is independent of DD, while RR is a function of the same summary coordinates as DD, and in Lemma 2.18 to write EF~a(d)\mathbb{E}\widetilde{F}^{(d)}_a (Step 3(d) below). The remark at the end of Section 2.6 that RR has full rank by Gaussianity is not used in the proof of Theorem 2.1. We now go through the steps of the source proof and record, for each, what it uses.

Step 1: the weights alone. Because Ti∈[τ,M]T_i\in[\tau,M] almost surely, DD is invertible and the empirical measure of DD is τ′\tau'-regular in the sense of Definition 2.2 for every τ′∈(0,min⁡{τ,M−1})\tau'\in(0,\min\{\tau,M^{-1}\}), deterministically: its support lies in [τ,M]⊂[−τ′−1,τ′−1][\tau,M]\subset[-\tau'^{-1},\tau'^{-1}] and it gives mass 0≤1−τ′0\le1-\tau' to {∣x∣≤τ′}\{|x|\le\tau'\}. (The endpoint τ′=τ\tau'=\tau is excluded: if μT\mu_T has an atom at τ\tau, the mass of {∣x∣≤τ}\{|x|\le\tau\} can exceed 1−τ1-\tau.) Thus Lemma 2.3 holds with probability one. Assumption 2 holds because μT([0,τ))=0\mu_T([0,\tau))=0. Assumption 3 holds almost surely by Lemma 4.3. Lemmas 2.4–2.6 are deterministic statements about τ′\tau'-regular measures. Since τ′\tau' may be taken arbitrarily small, the ball B(τ′−1)B(\tau'^{-1}) appearing in Theorem 2.12 may be taken to contain any prescribed compact set.

Step 2: the Gaussian block with fixed weights. The following results of the source involve only WW and DD: Proposition 2.7 (bulk law of WDWTWDW^{\mathsf T}), Proposition 2.8 (no eigenvalues of WDWTWDW^{\mathsf T} in EδE_\delta eventually, a modification of [1]), Theorem 2.9 and Lemmas 2.10–2.11 (deterministic equivalent for the Green function of the linearization [−zWWT−D−1]\begin{bmatrix}-z&W\\W^{\mathsf T}&-D^{-1}\end{bmatrix}), Theorem 2.12, and Lemma 2.13 (convergence of the finite-dd transform S(d)S^{(d)}). Their hypotheses on DD are those verified in Step 1 (Proposition 2.7 needs only the weak convergence in Assumption 3; Proposition 2.8 and Theorem 2.12 also use its Hausdorff part, Theorem 2.9 uses invertibility and τ′\tau'-regularity, and Lemmas 2.10–2.11 use τ′\tau'-regularity and the event of Proposition 2.8), and their only hypothesis linking DD and WW is independence. Fix a realization of the marks in the G\mathcal{G}-measurable probability-one event of Step 1 and Lemma 4.2. Along it the sequence (D(d))d(D^{(d)})_d is deterministic and satisfies these hypotheses, and (W(d))d(W^{(d)})_d is Gaussian and independent of it, so these results apply and their conclusions hold with conditional probability one; integrating over the marks (Fubini’s theorem), they hold almost surely. Theorem 2.12 allows the test vectors to be random provided they are independent of WW. We apply it with the qq columns of R~T\widetilde{R}^{\mathsf T}, where R~=[0q×(d−q) RT]\widetilde{R}=[0_{q\times(d-q)}\ R^{\mathsf T}]; these are G\mathcal{G}-measurable, hence independent of WW conditionally on G\mathcal{G}. Their norms are bounded almost surely, because ∥R∥F2=λd−1∑i∥Vi∥2→λE∥V∥2<∞\|R\|_F^2=\lambda d^{-1}\sum_i\|V_i\|^2\to\lambda\mathbb{E}\|V\|^2<\infty almost surely by Lemma 4.2 and (30).

Hypotheses of Theorem 2.9. Theorem 2.9 of the source is stated for a fixed τ>0\tau>0, an invertible diagonal DD whose empirical spectral measure is τ\tau-regular, and an independent Gaussian matrix with i.i.d. entries. Its conclusion is a bound Cd−1/4(Im⁡z)−5Cd^{-1/4}(\operatorname{Im}z)^{-5} on ∣⟨v,(G(z)−Π(z))u⟩∣|\langle v,(G(z)-\Pi(z))u\rangle| for each pair of deterministic unit vectors, with probability 1−O(n−100)1-O(n^{-100}), uniformly in ∣z∣≤τ−1|z|\le\tau^{-1}. Its proof follows [6], Section 5 and Lemma 6.2. The standing assumptions there [6], Section 2.1 are: the population matrix (here DD) is deterministic, and positive definite in the diagonal case; the noise matrix has independent centred entries of variance 1/N1/N, where NN is the dimension of the block carrying zz, and every moment of N\sqrt{N} times an entry is bounded; the ratio of the dimension of the population matrix to NN lies in [τ,τ−1][\tau,\tau^{-1}]; the population matrix has norm at most τ−1\tau^{-1}, and its empirical eigenvalue distribution gives mass at most 1−τ1-\tau to [0,τ][0,\tau]. The high-probability bounds there are uniform over all data satisfying these conditions with the same τ\tau [6, Definition 3.4]. Here the population dimension is mdm_d and N=d−qN=d-q; the letter MM keeps its meaning as the upper bound of the weights. Fix a realization of the marks as in Step 2 and put τ1=min⁡{τ′,c/2,1/(2c)}\tau_1=\min\{\tau',c/2,1/(2c)\}. For all large dd, DD is deterministic, diagonal and positive definite (the source’s extension to weights of either sign is not needed), ∥D∥≤M≤τ1−1\lVert D\rVert\le M\le\tau_1^{-1}, DD has no eigenvalue in [0,τ1][0,\tau_1], τ1≤md/(d−q)≤τ1−1\tau_1\le m_d/(d-q)\le\tau_1^{-1}, the matrix W\mathcal{W} is Gaussian and independent of the marks, and the test vectors used below are G\mathcal{G}-measurable.

Two points are not covered literally. First, the normalization. The source states Theorem 2.9 for a matrix with standard Gaussian entries, and Propositions 2.7–2.8 for n−1WDWTn^{-1}WDW^{\mathsf T} with standard Gaussian WW, but applies all three, in its Sections 2.9–2.10, to the matrix W\mathcal{W} of its Section 2.6. We read them in that normalization, in which the limiting measure is the one solving (2.11). The variance 1/(d−q)1/(d-q) is the one under which (2.11), with ratio md/(d−q)m_d/(d-q), is the correct equation; W\mathcal{W} has variance 1/d1/d, as in Section 2.6 of the source. Write κd=(d−q)/d\kappa_d=(d-q)/d and W=κd1/2W~W=\kappa_d^{1/2}\widetilde{W}. Then

[−zWWT−D−1]=[I00κd1/2I][−zW~W~T−(κdD)−1][I00κd1/2I].\begin{bmatrix} -z & \mathcal{W} \\ \mathcal{W}^{\mathsf T} & -D^{-1} \end{bmatrix} = \begin{bmatrix} I & 0 \\ 0 & \kappa_d^{1/2}I \end{bmatrix} \begin{bmatrix} -z & \widetilde{W} \\ \widetilde{W}^{\mathsf T} & -(\kappa_dD)^{-1} \end{bmatrix} \begin{bmatrix} I & 0 \\ 0 & \kappa_d^{1/2}I \end{bmatrix}.

and Theorem 2.9 applies to W~\widetilde{W} and κdD\kappa_dD, which is τ1\tau_1-regular for large dd because τ′<τ\tau'<\tau. The resulting deterministic equivalent for the original Green function has diagonal blocks S~(d)Id−q\widetilde{S}^{(d)}I_{d-q} and −D(I+κdS~(d)D)−1-D(I+\kappa_d\widetilde{S}^{(d)}D)^{-1}, where S~(d)\widetilde{S}^{(d)} solves (2.11) for the empirical law of the κdTi\kappa_dT_i. These laws converge weakly to μT\mu_T and are supported in [τ′/2,M][\tau'/2,M] for large dd, so Lemma 4.4, with τ′/2\tau'/2 in place of τ\tau, gives S~(d)→S\widetilde{S}^{(d)}\to S and κdS~(d)→S\kappa_d\widetilde{S}^{(d)}\to S locally uniformly on C+\mathbb{C}_+, the same limit as S(d)S^{(d)} in item (4) below. The extension to compact subsets of EδE_\delta is as in item (4). Since WDWT=W~(κdD)W~T\mathcal{W}D\mathcal{W}^{\mathsf T}=\widetilde{W}(\kappa_dD)\widetilde{W}^{\mathsf T}, the same rescaling applies to Propositions 2.7–2.8. This discrepancy, of relative order q/dq/d, is already present in the source, and it changes no limit below.

Second, the source’s Theorem 2.9 does without the lower bound on ∣1+m(z)σi∣\lvert1+m(z)\sigma_i\rvert and the stability condition assumed in [6 Section 3.4 and Definition 5.4]; the source replaces them by bounds in terms of Im⁡z\operatorname{Im}z which it takes from [1 (3.20)–(3.21)]. This modification is internal to the source’s proof and is not re-checked here. It is used only at spectral parameters with Im⁡z≥d−1/24\operatorname{Im}z\ge d^{-1/24} (Lemma 2.11) and ∣z∣\lvert z\rvert bounded above and below, which lie in the spectral domain of [6 Section 3.1] after decreasing τ1\tau_1.

Avoiding the modification for (ii)–(v). The outlier assertions do not need the unverified part of the proof of Theorem 2.9. Work in the scale of XdX_d, let ν\nu be the limiting measure of XdX_d, put [a′,b′]=[λc,a][a',b']=[\lambda c,a], and fix δ0∈(0,1)\delta_0\in(0,1) with a′−3δ0>sup⁡supp⁡νa'-3\delta_0>\sup\operatorname{supp}\nu. For (ii)–(v), Steps 2–5 need Lemma 2.11 and Theorem 2.12 only for zz with Re⁡z∈[a′−δ0,b′+δ0]\operatorname{Re}z\in[a'-\delta_0,b'+\delta_0] and 0≤Im⁡z0\le\operatorname{Im}z small (the lower half-plane follows from G(z‾)=G(z)∗G(\overline{z})=G(z)^*), so they use Theorem 2.9 only at points with Re⁡z∈[a′−δ0,b′+δ0]\operatorname{Re}z\in[a'-\delta_0,b'+\delta_0] and d−1/24≤Im⁡z≤1d^{-1/24}\le\operatorname{Im}z\le1. Let ϱd\varrho_d be the measure whose Stieltjes transform is S~(d)\widetilde{S}^{(d)}; it is the measure ϱ\varrho of [6 Definition 2.3] for Σ=κdD\Sigma=\kappa_dD and ratio md/(d−q)m_d/(d-q). By Lemma 4.5, applied to the empirical law of the κdTi\kappa_dT_i (support in [τ′/2,M][\tau'/2,M] for large dd, and converging weakly to μT\mu_T), dist⁡(Re⁡z,supp⁡ϱd)≥δ0\operatorname{dist}(\operatorname{Re}z,\operatorname{supp}\varrho_d)\ge\delta_0 at these points for all large dd. These points therefore lie in the domain D∘\mathbf{D}^{\circ} of [6 Section 3.3], for suitable constants. Theorem 3.16(i) of [6], the anisotropic local law outside the spectrum, assumes only [6 Assumption 2.1 and (2.9)], which were checked above with X=W~X=\widetilde{W} and Σ=κdD\Sigma=\kappa_dD. It assumes no regularity of the spectrum, no lower bound on ∣1+mσi∣\lvert1+m\sigma_i\rvert, and no stability condition. For fixed zz, D′>0D'>0 and deterministic unit vectors u,vu,v, it gives

∣⟨v,(G(z)−Π(z))u⟩∣≤Nεmax⁡{1,M}2Ψ(z),Ψ(z)≤2N1/2Im⁡z,\lvert\langle v,(G(z)-\Pi(z))u\rangle\rvert\le N^\varepsilon\max\{1,M\}^2\Psi(z),\qquad\Psi(z)\le\frac{2}{N^{1/2}\operatorname{Im}z},

with probability at least 1−N−D′1-N^{-D'}, for all large dd, uniformly in these zz; here GG and Π\Pi are written for (W~,κdD)(\widetilde{W},\kappa_dD), and the bound Ψ(z)≤2N−1/2(Im⁡z)−1\Psi(z)\le2N^{-1/2}(\operatorname{Im}z)^{-1} follows from Im⁡m≤∣m∣≤(Im⁡z)−1\operatorname{Im}m\le\lvert m\rvert\le(\operatorname{Im}z)^{-1}. For large dd this is smaller than the bound Cd−1/4(Im⁡z)−5Cd^{-1/4}(\operatorname{Im}z)^{-5} of Theorem 2.9. A union bound over a grid of mesh d−3d^{-3} makes it simultaneous in zz, because GG and Π\Pi are Lipschitz in zz there with constants polynomial in dd on the event ∥W~∥≤C0\lVert\widetilde{W}\rVert\le C_0, whose complement has probability at most e−de^{-d} for C0C_0 large. Lemmas 2.10–2.11 and Theorem 2.12, restricted to a complex neighbourhood of [a′,b′][a',b'], then hold as stated; their proofs still use Proposition 2.8. The modified Theorem 2.9 therefore enters only the alternative argument for (i) in item (2) of the next paragraph.

Aspect ratio. The source states Theorem 2.1 with n/d=ϕn/d=\phi; here md/d→cm_d/d\to c. The ratio enters the proof in the following places. (1) Proposition 2.8 is first stated for n/d→ϕn/d\to\phi; in its proof the function ψd\psi_d is formed with the finite-dd ratio, and the uniform convergence ψd→ψ\psi_d\to\psi, ψd′→ψ′\psi'_d\to\psi' on [a,b][a,b] persists because $\phi_d\to c. (2) Proposition 2.7 is quoted by the source from Silverstein and Bai. Independently of how that theorem treats the ratio, its conclusion here also follows from results already listed: applying Theorem 2.9 to the d−qd-q coordinate vectors of the upper block, a union bound and the Borel–Cantelli lemma show that, for each fixed z∈C+z\in\mathbb{C}^{+}, the normalized trace of (WDWT−z)−1(WDW^{\mathsf{T}}-z)^{-1} minus S~(d)(z)\widetilde{S}^{(d)}(z) (in the normalization above) tends to zero almost surely; S~(d)(z)\widetilde{S}^{(d)}(z) converges to the limiting transform, as shown above, and convergence of Stieltjes transforms on a countable dense subset of C+\mathbb{C}^{+} to the Stieltjes transform of a probability measure gives almost sure weak convergence. (3) Theorem 2.9 and Lemmas 2.4–2.6 and 2.10–2.11 are finite-dd statements written with the finite-dd ratio of (2.11), which varies with dd already in the source. (4) Lemma 2.13: by Lemma 4.4(b), applied to the empirical weight laws (which converge weakly by Lemma 4.3) and to ϕd=md/(d−q)→c\phi_d=m_d/(d-q)\to c, the finite-dd transforms converge to the limiting one locally uniformly on C+\mathbb{C}^{+}; Lemma 4.4(a) also identifies the source’s S(d)S^{(d)} as the unique solution of (2.11) in C+\mathbb{C}^{+}. The extension to compact subsets of EδE_\delta is then as in the source, through Proposition 2.8 and Lemma 2.10. (5) Lemmas 2.16–2.18 contain the factor n/dn/d; here it is cd→cc_d\to c, see Step 3.

Step 3: the reduced matrix. Theorem 2.14 states that F(d)(z)=RTD(I+S(d)(z)D)−1RF^{(d)}(z)=R^{\mathsf{T}}D(I+S^{(d)}(z)D)^{-1}R converges to its deterministic limit uniformly on Eδ,K,ηE_{\delta,K,\eta}. Its proof uses four ingredients.

(a) Lemma 2.15 compares D(I+S(d)D)−1D(I+S^{(d)}D)^{-1} with a cut-off version; it uses only Step 1, Lemma 2.13 and Lemmas 2.5–2.6.

(b) Lemma 2.16 combines Lemma 2.15 with almost sure boundedness of

1d∑ℓ(dRℓi)2(dRℓj)2=λ2d∑ℓVℓi2Vℓj2,\frac{1}{d}\sum_{\ell}(\sqrt{d}R_{\ell i})^{2}(\sqrt{d}R_{\ell j})^{2} =\frac{\lambda^{2}}{d}\sum_{\ell}V_{\ell i}^{2}V_{\ell j}^{2},

which holds by Lemma 4.2 and (4.1).

(c) Lemma 2.17 applies the law of large numbers, at fixed zz, to i.i.d. summands λθδ(1+S(z)Tℓ)Tℓ(1+S(z)Tℓ)−1VℓVℓT\lambda\theta_\delta(1+S(z)T_\ell)T_\ell(1+S(z)T_\ell)^{-1}V_\ell V_\ell^{\mathsf{T}}, where SS is the deterministic limit and θδ\theta_\delta is the smooth cut-off; the cut-off bounds the scalar factor, so all moments are finite by (4.1) and Lemma 4.2 applies. Equicontinuity in zz uses the same bound times d−1∑ℓ∥Vℓ∥d^{-1}\sum_\ell\|V_\ell\|, which converges almost surely. A countable dense set of zz and the Arzelà–Ascoli theorem give local uniformity, as in the source.

(d) Lemmas 2.18–2.19 identify the mean. Because the summands are i.i.d. with law determined by μ\mu,

EF~δ(d)(z)=cdλE[θδ(1+S(z)T)T1+S(z)TVVT](38)\mathbb{E}\widetilde{F}^{(d)}_\delta(z)=c_d\lambda\mathbb{E}\left[\theta_\delta(1+S(z)T)\frac{T}{1+S(z)T}VV^{\mathsf{T}}\right] \tag*{(38)}

which converges to cλE[⋯ ]c\lambda\mathbb{E}[\cdots]; since μ\mu does not depend on dd, Lemma 2.19 is not needed. The source uses Gaussianity of gg only to rewrite this expectation in terms of gg and GG. For the removal of the cut-off as δ→0\delta\to0, note that for positive weights ∣1+S(z)t∣\lvert1+S(z)t\rvert is bounded below by some κ>0\kappa>0 under supp⁡μ\operatorname{supp}\mu and zz in a compact subset of EδE_\delta: for nonreal zz, Im⁡(1+S(z)t)=tIm⁡S(z)≠0\operatorname{Im}(1+S(z)t)=t\operatorname{Im}S(z)\ne0; for real zz, 1+S(z)t≠01+S(z)t\ne0 by the Silverstein–Choi characterization used in the paragraph Denominators below (which involves only the limiting equation); continuity and compactness give κ\kappa. For δ<κ\delta<\kappa the cut-off is identically one, so the truncated and untruncated limits coincide.

Therefore F(d)→F~F^{(d)}\to\widetilde{F} uniformly on Eδ,K,ηE_{\delta,K,\eta} almost surely, where

F~(z)=cλE[T1+S(z)TVVT].(39)\widetilde{F}(z)=c\lambda\mathbb{E}\left[\frac{T}{1+S(z)T}VV^{\mathsf{T}}\right]. \tag*{(39)}

Step 4: linearization and Schur complement. Since DD is invertible, Section 2.10 of the source applies verbatim: rotating by L⊕L⊥L\oplus L^\perp, the linearization of XdX_d becomes

[−zIqR~R~TN(z)]\begin{bmatrix} -zI_q & \widetilde{R}\\ \widetilde{R}^{\mathsf{T}} & \mathcal{N}(z) \end{bmatrix}

with

N(z)=[−zId−qWWT−D−1].\mathcal{N}(z)= \begin{bmatrix} -zI_{d-q} & W\\ W^{\mathsf{T}} & -D^{-1} \end{bmatrix}.

By Step 2, N(z)\mathcal{N}(z) is invertible on EδE_\delta for all large dd. Put Δd(z)=zIq+R~N(z)−1R~T\Delta_d(z)=zI_q+\widetilde{R}\mathcal{N}(z)^{-1}\widetilde{R}^{\mathsf{T}}. Since (Xd−z)−1(X_d-z)^{-1} is the upper-left block of the inverse linearization, the Schur complement formula gives

LT(Xd−z)−1L=−Δd(z)−1,(40)L^{\mathsf{T}}(X_d-z)^{-1}L=-\Delta_d(z)^{-1}, \tag*{(40)}

and the eigenvalues of XdX_d in Eδ∩RE_\delta\cap\mathbb{R} are the zeros of det⁡Δd\det\Delta_d, with multiplicity equal to the order of the zero. By Step 2 and the bounded norms of the columns of R~T\widetilde{R}^{\mathsf{T}}, Δd(z)−(zIq−F(d)(z))→0\Delta_d(z)-(zI_q-F^{(d)}(z))\to0 uniformly, and by Step 3, Δd→zIq−F~\Delta_d\to zI_q-\widetilde{F} uniformly on Eδ,K,ηE_{\delta,K,\eta}, almost surely. Since these functions are holomorphic, their derivatives converge uniformly on slightly smaller sets by the Cauchy estimates. Hurwitz’s theorem gives the root counts, exactly as in the source.

Step 5: eigenvectors. Sections 2.11–2.12 of the source use only the linearization identity (2.51)–(2.52), the invertibility of N\mathcal{N}, and the convergence of Δd\Delta_d and Δd′\Delta'_d from Step 4. They give, for an eigenvalue zd→z∗≠0z_d \to z_\ast\ne0 of X~d\widetilde{X}_d with unit eigenvector udu_d, that every limit point uu of LTudL^T u_d satisfies (z∗−F~(z∗))u=0(z_\ast-\widetilde{F}(z_\ast))u=0 and ∥u∥2−uTF~′(z∗)u=1\lVert u\rVert^2-u^T\widetilde{F}'(z_\ast)u=1. In fact the computation there gives, for vd=LTudv_d=L^T u_d, the exact identities Δd(zd)vd=0\Delta_d(z_d)v_d=0 and vdTΔd′(zd)vd=1v_d^T\Delta'_d(z_d)v_d=1: with N\mathcal{N} real symmetric for real zz, Δd′=Iq+N−1(Id−q⊕0)N−1R~T\Delta'_d=I_q+\mathcal{N}^{-1}(I_{d-q}\oplus0)\mathcal{N}^{-1}\widetilde{R}^T, and the second term evaluated at vdv_d is the squared norm of the component of udu_d orthogonal to Ran⁡L\operatorname{Ran}L. Together with the uniform convergence of Δd\Delta_d and Δd′\Delta'_d from Step 4 these identities give the limits in (v) directly.

Rescaling. Let S(z)S(z) be the limiting transform of XdX_d, so S(λcx)=sc(x)/(λc)S(\lambda c x)=s_c(x)/(\lambda c). The equation 1+zS=cE[ST/(1+ST)]1+zS=c\mathbb{E}[ST/(1+ST)] for XdX_d becomes (32), and (4.10) gives $\widetilde{F}(\lambda c x)=\lambda cF_c(x)$, hence $\widetilde{F}'(\lambda c x)=F'_c(x)$. Define $B_d(x)=(\lambda c_d)^{-1}\Delta_d(\lambda c_d x)$. Then (4.11) becomes {xref}(4.5)`, and, since cd→cc_d\to c, Bd→BcB_d\to B_c together with the derivative, uniformly on a thin complex neighbourhood Ω\Omega of [a,b][a,b] contained in the rescaled set Eδ,K,ηE_{\delta,K,\eta}. This proves (iv); (iii) and (v) follow from Steps 4–5, and (i) from Section 2.9 of the source, which uses only Proposition 2.7 and Cauchy interlacing for a perturbation of rank at most 2q2q.

Denominators. For (ii), the Silverstein–Choi characterization of the complement of the support [11 Theorem 4.1], applied to the bulk measure of XdX_d, which solves (2.11) with weight law μT\mu_T, shows that −1/S(z)∉supp⁡μT-1/S(z)\notin\operatorname{supp}\mu_T for real zz outside the support; this characterization is also the one used in the proof of Proposition 2.8. Equivalently λc+sc(x)t≠0\lambda c+s_c(x)t\ne0 for x>λ+,cx>\lambda_{+,c} and t∈supp⁡μTt\in\operatorname{supp}\mu_T. Continuity of scs_c on KK and compactness of K×supp⁡μTK\times\operatorname{supp}\mu_T give a positive infimum, and real analyticity follows by differentiating under the expectation. □

Remark 4.6 (Where the inverse weight matrix enters). In the source proof the diagonal inverse appears only in the lower-right block of the linearization, in the Schur complement producing the characteristic determinant, in the relation (2.52) between the eigenvector and the linearization vector, and in the differentiation of the Schur complement giving the norm identity. After (26), each occurrence is an inverse of D+,dD_{+,d} and is bounded by τ−1\tau^{-1}. The bulk theorem does not require the inverse and already permits a zero atom. No Moore–Penrose substitution is made. The rotation-invariance argument of Section 2.13 of the source, which uses the Gaussian law of gg, is not used; its root bound is replaced by the monotonicity argument of Section 7.

Bulk law, exact scaling, and the zero atom

Under the active law (27), set

T=hb(g),V=mb+λ−1/2g,c=ϕπ.(41)T=h_b(g), \qquad V=m_b+\lambda^{-1/2}g, \qquad c=\phi\pi. \tag*{(41)}

This mark law satisfies the hypotheses of Proposition 4.1: T∈[τ,M]T\in[\tau,M], and E+∥V∥2≤τ−1E∥mb+λ−1/2g∥2<∞\mathbb{E}_+\lVert V\rVert^2\le\tau^{-1}\mathbb{E}\lVert m_b+\lambda^{-1/2}g\rVert^2<\infty. Let s+s_+ solve

1+xs+(x)=cE+[s+(x)Tλc+s+(x)T].(42)1+xs_+(x)=c\mathbb{E}_+\left[\frac{s_+(x)T}{\lambda c+s_+(x)T}\right]. \tag*{(42)}

Define

S(z)=1πs+(zπ).(43)S(z)=\frac{1}{\pi}s_+\left(\frac{z}{\pi}\right). \tag*{(43)}

Since

E+[f(b,g)]=1π∑bpbE[1{hb(g)>0}f(b,g)],(44)\mathbb{E}_+[f(b,g)]=\frac{1}{\pi}\sum_b p_b\mathbb{E}\left[\mathbf{1}_{\{h_b(g)>0\}}f(b,g)\right], \tag*{(44)}

substitution in (42) yields (8). Inactive terms vanish because the numerator is zero; no inverse of zero is used.

If ν+\nu_+ is the active bulk measure, then

νG=(x↦πx)#ν+,supp⁡νG=πsupp⁡ν+,λ+(G)=πλ+,+.(45)\nu_G=(x\mapsto\pi x)_\#\nu_+,\qquad\operatorname{supp}\nu_G=\pi\operatorname{supp}\nu_+,\qquad\lambda_+(G)=\pi\lambda_{+,+}. \tag*{(45)}

This exact pushforward supplies the right-edge statement without invoking continuity of support under weak convergence.

For the effective matrix,

F+(x)=λcE+[Tλc+s+(x)TVVT].(46)F_{+}(x)=\lambda c E_{+}\left[\frac{T}{\lambda c+s_{+}(x)T}VV^{\mathsf T}\right]. \tag*{(46)}

Using c=ϕπc=\phi\pi and (43),

FG(z)=πF+(zπ),(47)F_G(z)=\pi F_{+}\left(\frac{z}{\pi}\right), \tag*{(47)}

which expands exactly to (11). Hence

det⁡BG(z)=πqdet⁡[zπIq−F+(zπ)].(48)\det B_G(z)=\pi^q\det\left[\frac{z}{\pi}I_q-F_{+}\left(\frac{z}{\pi}\right)\right]. \tag*{(48)}

Roots and analytic multiplicities correspond bijectively under z=πxz=\pi x.

Lemma 5.1 (Zero mass). The limiting measure satisfies (12).

Proof. The rank theorem proved in § 6 gives the lower bound

νG({0})≥(1−ϕπ)+.(49)\nu_G(\{0\})\ge(1-\phi\pi)_+. \tag*{(49)}

Alternatively, the same lower bound follows from (45) and the standard active covariance rank deficit. Let m0=νG({0})m_0=\nu_G(\{0\}). For t↓0t\downarrow0,

tSG(−t)⟶m0.(50)tS_G(-t)\longrightarrow m_0. \tag*{(50)}

If m0>0m_0>0, then SG(−t)→+∞S_G(-t)\to+\infty. Dominated convergence in (8) gives

1−m0=ϕ∑bpbP{hb(g)>0}=ϕπ.(51)1-m_0=\phi\sum_b p_b\mathbb{P}\{h_b(g)>0\}=\phi\pi. \tag*{(51)}

Thus m0=1−ϕπm_0=1-\phi\pi when ϕπ<1\phi\pi<1. If ϕπ≥1\phi\pi\ge1, a positive m0m_0 would imply 1−m0=ϕπ≥11-m_0=\phi\pi\ge1, which is impossible. Therefore m0=(1−ϕπ)+m_0=(1-\phi\pi)_+. □

Exact rank and kernel

Lemma 6.1 (General position of active columns). Conditioned on activity, each YℓY_\ell has an absolutely continuous law on Rd\mathbb{R}^d. Independent active columns are in general linear position almost surely.

Proof. Under (27), gg has a density on its active set, and ww remains a full standard Gaussian. For each class bb, the affine orthogonal map

(g,w)↦Ld(mb+λ−1/2g)+λ−1/2Ld⊥w(52)(g,w)\mapsto L_d(m_b+\lambda^{-1/2}g)+\lambda^{-1/2}L_d^\perp w \tag*{(52)}

has nonsingular linear part. Its image law is absolutely continuous on Rd\mathbb{R}^d. Inductively, conditioned on the first r<dr<d active columns, their span is a proper linear subspace and hence has Lebesgue measure zero. The next active column avoids it almost surely. □

Therefore

rank⁡A+,d=min⁡(d,Nd)a.s.(53)\operatorname{rank}A_{+,d}=\min(d,N_d)\quad\text{a.s.} \tag*{(53)}

Since D+,d1/2D_{+,d}^{1/2} is invertible,

rank⁡Md=rank⁡(A+,dD+,d1/2)=min⁡(d,Nd)(54)\operatorname{rank}M_d=\operatorname{rank}(A_{+,d}D_{+,d}^{1/2})=\min(d,N_d) \tag*{(54)}

almost surely. Equation (22) follows from (29). The exact zeros are the finite-dimensional realization of the atom (12); they are not positive outliers.

Analyticity and root multiplicity

On a compact set to the right of the active bulk, Proposition 4.1(ii) supplies

ess inf⁡T∣λc+s+(x)T∣>0.(55)\underset{T}{\operatorname{ess\,inf}}\lvert\lambda c+s_+(x)T\rvert>0. \tag*{(55)}

Equations (43) and (47) imply (13). Thus FGF_G is real analytic on (λ+,∞)(\lambda_+,\infty).

For real z>λ+z>\lambda_+,

SG′(z)=∫νG(dx)(x−z)2>0.(56)S'_G(z)=\int\frac{\nu_G(dx)}{(x-z)^2}>0. \tag*{(56)}

Differentiating (11),

FG′(z)=−λϕSG′(z)∑bpbE[hb(g)2vb(g)vb(g)T[λϕ+SG(z)hb(g)]2]⪯0.(57)F'_G(z)=-\lambda\phi S'_G(z)\sum_b p_b\mathbb{E}\left[\frac{h_b(g)^2v_b(g)v_b(g)^{\mathsf T}}{[\lambda\phi+S_G(z)h_b(g)]^2}\right]\preceq0. \tag*{(57)}

Hence

BG′(z)=Iq−FG′(z)⪰Iq.(58)B'_G(z)=I_q-F'_G(z)\succeq I_q. \tag*{(58)}

Every ordered eigenvalue of the Hermitian matrix BG(z)B_G(z) is strictly increasing, so each can cross zero at most once and the sum of root multiplicities is at most qq.

Let z∗z_* be a root and K∗=Ker⁡BG(z∗)K_*=\operatorname{Ker}B_G(z_*), dim⁡K∗=m\dim K_*=m. In the orthogonal decomposition K∗⊕K∗⊥K_*\oplus K_*^\perp,

BG(z∗+t)=[tJ∗+O(t2)tC+O(t2)tCT+O(t2)B22+O(t)],(59)B_G(z_*+t)= \begin{bmatrix} tJ_*+O(t^2) & tC+O(t^2)\\ tC^{\mathsf T}+O(t^2) & B_{22}+O(t) \end{bmatrix}, \tag*{(59)}

where B22B_{22} is invertible and

J∗=P∗BG′(z∗)P∗∣K∗≻0.(60)J_*=P_*B'_G(z_*)P_*|_{K_*}\succ0. \tag*{(60)}

The Schur complement on K∗K_* equals tJ∗+O(t2)tJ_*+O(t^2), and therefore

det⁡BG(z∗+t)=det⁡(B22)tmdet⁡(J∗)+O(tm+1).(61)\det B_G(z_*+t)=\det(B_{22})t^m\det(J_*)+O(t^{m+1}). \tag*{(61)}

This proves (15).

Random active size and empirical outlier counts

Lemma 8.1 (Random-index transfer). Let rd,m∈[0,1]r_{d,m}\in[0,1] be a failure probability for a model with mm active columns. Suppose

rd,md→0(62)r_{d,m_d}\to0 \tag*{(62)}

for every deterministic sequence md/d→cm_d/d\to c. If Nd/d→cN_d/d\to c in probability, then

Erd,Nd→0.(63)\mathbb{E}r_{d,N_d}\to0. \tag*{(63)}

Proof. Choose ηd↓0\eta_d\downarrow0 such that

P{∣Nd/d−c∣>ηd}→0.(64)\mathbb{P}\{|N_d/d-c|>\eta_d\}\to0. \tag*{(64)}

If

sup⁡∣m/d−c∣≤ηdrd,m(65)\sup_{\lvert m/d-c\rvert\le\eta_d}r_{d,m} \tag*{(65)}

did not tend to zero, a subsequence and a choice md/d→cm_d/d\to c would contradict the deterministic-sequence hypothesis. Splitting the expectation on the typical and atypical events gives (63). □

For each deterministic md/d→c=ϕπm_d/d\to c=\phi\pi, Proposition 4.1 gives the exact active eigenvalue count on compact intervals separated from the active bulk; almost sure eventual statements imply that the corresponding failure probabilities tend to zero. Because

Md=αdM^d,αd=Ndn→π,(66)M_d=\alpha_d\widehat{M}_d,\qquad\alpha_d=\frac{N_d}{n}\to\pi, \tag*{(66)}

one has

#{σ(Md)∩[a,b]}=#{σ(M^d)∩[aαd,bαd]}.(67)\#\{\sigma(M_d) \cap[a,b]\}=\#\{\sigma(\widehat{M}_d) \cap[\frac{a}{\alpha_d},\frac{b}{\alpha_d}]\}. \tag*{(67)}

The active endpoints converge to a/πa/\pi and b/πb/\pi. Uniform determinant convergence in a small endpoint neighbourhood permits these converging endpoints. Equation (48) identifies the active root count with the target root count. The random-index lemma removes conditioning and proves (14).

No spurious right outliers and existence

The operator norm is tight because

0⪯Md⪯MnAdAdT.(68)0 \preceq M_d \preceq\frac{M}{n}A_dA_d^{\mathsf{T}}. \tag*{(68)}

Writing Ad=Zd+MdA_d=Z_d+M_d, the Gaussian norm satisfies ∥Zd∥/n=OP(1)\lVert Z_d\rVert/\sqrt{n}=O_{\mathrm{P}}(1), while

∥Md∥n≤max⁡b∥μb(d)∥=O(1).(69)\frac{\lVert M_d\rVert}{\sqrt{n}}\leq\max_b\lVert\mu_b^{(d)}\rVert=O(1). \tag*{(69)}

Thus ∥Md∥=OP(1)\lVert M_d\rVert=O_{\mathrm{P}}(1).

Also FG(z)F_G(z) remains bounded as z→+∞z\to+\infty, while zIqzI_q diverges, so effective roots lie in a bounded interval. Fix ϵ>0\epsilon>0 and a large deterministic KK. The right roots in [λ++ϵ,K][\lambda_++\epsilon,K] are finite. Remove small disjoint neighbourhoods of these roots. The complement is a finite union of compact root-free intervals, and the count theorem gives no empirical eigenvalues there with probability tending to one. This proves the no-spurious direction. Applying the count theorem to an isolating interval, and then to nested isolating intervals, proves existence and convergence at each effective root.

Simple roots and spectral projections

For the active model, Proposition 4.1(v) gives, near an isolated root x∗=z∗/πx_* = z_*/\pi,

∥(xdI−F+(xd))rd∥=oP(1),(70)\lVert(x_dI-F_+(x_d))r_d\rVert=o_{\mathrm{P}}(1), \tag*{(70)}
∣rdT[I−F+′(xd)]rd−1∣=oP(1).(71)\left|r_d^{\mathsf{T}}[I-F_+'(x_d)]r_d-1\right|=o_{\mathrm{P}}(1). \tag*{(71)}

Multiplication by αd\alpha_d changes eigenvalues but not eigenvectors. From (47),

FG′(z)=F+′(zπ).(72)F_G'(z)=F_+'\left(\frac{z}{\pi}\right). \tag*{(72)}

Local uniformity and αd→π\alpha_d\to\pi yield (16)–(17). At a simple root, (15) makes the kernel one-dimensional, and (17) fixes the norm, proving (18).

For a repeated root, choose a positively oriented contour Γ\Gamma enclosing z∗z_* and no other effective root or bulk point, of radius smaller than the strip width in Proposition 4.1(iv). The active compressed resolvent identity (34), the uniform convergence Bd→BcB_d\to B_c, and the scaling (47) give

sup⁡z∈Γ∥LdT(Md−zI)−1Ld+BG(z)−1∥→P0.(73)\sup_{z\in\Gamma}\left\lVert L_d^{\mathsf{T}}(M_d-zI)^{-1}L_d+B_G(z)^{-1}\right\rVert\xrightarrow{\mathrm{P}} 0. \tag*{(73)}

The full-space spectral projection is

Πd=−12πi∮Γ(Md−zI)−1 dz.(74)\Pi_d=-\frac{1}{2\pi i}\oint_\Gamma(M_d-zI)^{-1}\,dz. \tag*{(74)}

Thus

LdTΠdLd→P12πi∮ΓBG(z)−1 dz.(75)L_d^{\mathsf{T}}\Pi_dL_d\xrightarrow{\mathrm{P}}\frac{1}{2\pi i}\oint_\Gamma B_G(z)^{-1}\,dz. \tag*{(75)}

The block expansion near z∗z_* gives

BG(z∗+t)−1=1tP∗J∗−1P∗+O(1).(76)B_G(z_*+t)^{-1}=\frac{1}{t}P_*J_*^{-1}P_*+O(1). \tag*{(76)}

The residue is P∗J∗−1P∗P_*J_*^{-1}P_*, proving (20). If UdU_d is any orthonormal basis of Ran⁡Πd\operatorname{Ran}\Pi_d and Rd=LdTUdR_d=L_d^{\mathsf{T}}U_d, then

RdRdT→PP∗J∗−1P∗.(77)R_dR_d^{\mathsf{T}}\xrightarrow{\mathrm{P}}P_*J_*^{-1}P_*. \tag*{(77)}

Only the invariant subspace is canonical; individual eigenvectors may rotate.

ReLU gates

Corollary 11.1. Assume

hb(g)=ab(g)1{ℓb(g)>0},(78)h_b(g)=a_b(g)\mathbf{1}_{\{\ell_b(g)>0\}}, \tag*{(78)}

where ℓb\ell_b is a nonconstant affine functional and

τ≤ab(g)≤Mon {ℓb(g)>0}.(79)\tau\le a_b(g) \le M \quad\text{on } \{\ell_b(g)>0\}. \tag*{(79)}

Then Theorem 2.1 applies.

Proof. A nonconstant affine functional of a standard Gaussian is a nondegenerate one-dimensional Gaussian, so its positive half-space has probability strictly between zero and one. The weights belong to {0}∪[τ,M]\{0\} \cup[\tau,M] almost surely. □

For the pure gate hb(g)=1{ℓb(g)>0}h_b(g)=\mathbf{1}_{\{\ell_b(g)>0\}},

Md=1n∑ℓ=1n1{ℓb(gℓ)>0}YℓYℓT(80)M_d=\frac{1}{n}\sum_{\ell=1}^{n}\mathbf{1}_{\{\ell_b(g_\ell)>0\}}Y_\ell Y_\ell^{\mathsf T} \tag*{(80)}

is a half-space-gated empirical second moment. For a one-unit ReLU model

fw(Y)=amax⁡(wTY,0),(81)f_w(Y)=a\max(w^{\mathsf T}Y,0), \tag*{(81)}

at a fixed parameter for which no sample lies at the kink,

∇wfw(Y)=a1{wTY>0}Y.(82)\nabla_w f_w(Y)=a\mathbf{1}_{\{w^{\mathsf T}Y>0\}}Y. \tag*{(82)}

The ww–ww Jacobian-Gram or Gauss–Newton block is therefore

a2n∑ℓ=1n1{wTYℓ>0}YℓYℓT.(83)\frac{a^2}{n}\sum_{\ell=1}^{n}\mathbf{1}_{\{w^{\mathsf T}Y_\ell>0\}}Y_\ell Y_\ell^{\mathsf T}. \tag*{(83)}

For squared loss and fixed deterministic ww, Gaussian data avoid the kink almost surely, the activation pattern is locally constant, and the classical Hessian block equals this matrix. This does not establish a trajectory-uniform theorem, cross-neuron blocks, or a generalized Hessian at a kink. Hard ReLU gates thus create a zero atom, not additional positive outliers.

Examples and numerical illustration

Independent Bernoulli thinning

12.1. Independent Bernoulli thinning. If h∼Bernoulli⁡(π)h \sim\operatorname{Bernoulli}(\pi) independently of the columns, then

Md=Ndn(1NdA+A+T),Ndd→c=φπ.(84)M_d=\frac{N_d}{n}\left(\frac{1}{N_d}A_+A_+^{\mathsf T}\right),\qquad\frac{N_d}{d}\to c=\varphi\pi. \tag*{(84)}

The continuous bulk edges are

λ±=(1±c)2λϕ=πλ(1±1c)2,(85)\lambda_{\pm}=\frac{(1\pm\sqrt{c})^2}{\lambda\phi}=\frac{\pi}{\lambda}\left(1\pm\frac{1}{\sqrt{c}}\right)^2, \tag*{(85)}

and the zero mass is (1−c)+(1-c)_+. Equation (8) becomes

zS(z)2+(1+λϕz−φπ)S(z)+λϕ=0.(86)zS(z)^2+(1+\lambda\phi z-\varphi\pi)S(z)+\lambda\phi=0. \tag*{(86)}

Centred pure gate

12.2. Centred pure gate. Let Y=λ−1/2ZY=\lambda^{-1/2}Z with Z∼N(0,Id)Z\sim N(0,I_d) and h=1{⟨ℓ,Z⟩>0}h=\mathbf{1}_{\{\langle\ell,Z\rangle>0\}}. Conditional on ZZTZZ^{\mathsf T}, the only Gaussian preimages are ZZ and −Z-Z, with equal probability, and the gate takes complementary values. Therefore

P{h=1∣YYT}=12,(87)\mathbb{P}\{h=1\mid YY^{\mathsf T}\}=\frac{1}{2}, \tag*{(87)}

so hh is independent of YYTYY^{\mathsf T}. The model is exactly independent Bernoulli thinning with π=1/2\pi=1/2.

Noncentred scalar gate

Let g∼N(0,1)g \sim N(0,1),

h(g)=1{a+bg>0},b≠0,(88)h(g)=\mathbf{1}_{\{a+bg>0\}}, \qquad b\ne0, \tag*{(88)}

and define

δ=a∣b∣,sb=sign⁡(b),π=Φ(δ),κ=φ(δ)Φ(δ).(89)\delta=\frac{a}{|b|}, \qquad s_b=\operatorname{sign}(b), \qquad\pi=\Phi(\delta), \qquad\kappa=\frac{\varphi(\delta)}{\Phi(\delta)}. \tag*{(89)}

Then

E[g∣h=1]=sbκ,E[g2∣h=1]=1−δκ.(90)\mathbb{E}[g\mid h=1]=s_b\kappa, \qquad\mathbb{E}[g^2\mid h=1]=1-\delta\kappa. \tag*{(90)}

For v=m+λ−1/2gv=m+\lambda^{-1/2}g, set

c=φπ,β=λE[v2∣h=1]=λm2+2λ msbκ+1−δκ.(91)c=\varphi\pi, \qquad\beta=\lambda\mathbb{E}[v^2\mid h=1]=\lambda m^2+2\sqrt{\lambda}\,ms_b\kappa+1-\delta\kappa. \tag*{(91)}

The scalar effective equation gives the standard spiked-covariance threshold

β>1+1c(92)\beta>1+\frac{1}{\sqrt{c}} \tag*{(92)}

and, above threshold,

zout=πλ β(1+1c(β−1)).(93)z_{\mathrm{out}}=\frac{\pi}{\lambda}\,\beta\left(1+\frac{1}{c(\beta-1)}\right). \tag*{(93)}

Fixed-epsilon comparison

For comparison only, define

Dϵ=D+ϵP0,P0=diag⁡(1{Dℓℓ=0}).(94)D_\epsilon=D+\epsilon P_0, \qquad P_0=\operatorname{diag}(\mathbf{1}_{\{D_{\ell\ell}=0\}}). \tag*{(94)}

Then

Md,ϵ−Md=ϵnAdP0AdT(95)M_{d,\epsilon}-M_d=\frac{\epsilon}{n}A_dP_0A_d^{\mathsf T} \tag*{(95)}

and

∥Md,ϵ−Md∥≤ϵ∥Ad∥2n=OP(ϵ).(96)\lVert M_{d,\epsilon}-M_d\rVert\leq\epsilon\frac{\lVert A_d\rVert^2}{n}=O_{\mathbb{P}}(\epsilon). \tag*{(96)}

This perturbation estimate does not, by itself, prove support-edge or root-multiplicity stability. The exact active reduction avoids that missing limit interchange.

Numerical illustration

Figure 1 shows the noncentred scalar gate. The active conditional moment predicts an outlier at 4.6653, separated from the bulk edge 2. Direct simulations approach the prediction while the associated eigenvectors retain large overlap with the summary direction. Representative values from the same computations are listed below.

Largest eigenvalue for the gate $h(g)=\mathbf{1}_{\{g>0\}}$

Figure 1. Largest eigenvalue for the gate h(g)=1{g>0}h(g) = \mathbf{1}_{\{g>0\}} with ϕ=2\phi= 2, λ=1\lambda= 1, and summary mean m=2m = 2. The effective determinant predicts zout=4.6653z_{\mathrm{out}} = 4.6653 and the bulk edge is 2.

QuantityValue
Bernoulli model rank254 observed, 254 predicted
Bernoulli zero mass0.20625 observed, 0.2 limiting
Bernoulli top eigenvalue1.1641 observed, 1.1963 edge
Stieltjes residual2.39×10−162.39 \times10^{-16}
Centred gate top eigenvalue1.9728, edge 2
Noncentred predicted outlier4.6653
Noncentred simulations4.48–4.87
Determinant residual1.62×10−101.62 \times10^{-10}
Regularization identity error<1.1×10−17<1.1 \times10^{-17}
Two-construction operator difference1.93×10−151.93 \times10^{-15}

Table 2.

A two-class check with non-constant weights

The following check, added in the present revision, is numerical evidence only. Take k=2k = 2, C=0C = 0, q=2q = 2, p1=p2=1/2p_1 = p_2 = 1/2, λ=1\lambda= 1, ϕ=1.2\phi= 1.2 and G=4I2\sqrt{G} = 4I_2, so that m1=(4,0)m_1 = (4, 0) and m2=(0,4)m_2 = (0, 4). Let

h1(g)=a1(g1)1{g2>−0.3},h2(g)=a2(g1)1{g2<0.5},h_1(g) = a_1(g_1)\mathbf{1}_{\{g_2>-0.3\}}, \qquad h_2(g) = a_2(g_1)\mathbf{1}_{\{g_2<0.5\}},

where a1a_1 equals 0.5, 1, 1.5 on (−∞,−0.5)(-\infty, -0.5), [−0.5,0.7)[-0.5, 0.7), [0.7,∞)[0.7, \infty), and a2a_2 equals 0.6, 1.2 on (−∞,0](-\infty, 0], (0,∞)(0, \infty). This is a gate of the form of § 11 with τ=0.5\tau= 0.5 and M=1.5M = 1.5. Here π=0.6547\pi= 0.6547, so νG({0})=1−ϕπ=0.2144\nu_G(\{0\}) = 1 - \phi\pi= 0.2144. Because the weights are piecewise constant, every expectation in (8) and (11) is a finite sum of truncated Gaussian moments. Inverting the map SG↦zS_G \mapsto z defined by (8) gives λ+(G)=3.0976\lambda_+(G) = 3.0976. Bracketing the increasing eigenvalues of BGB_G gives two simple roots, z1=7.8129z_1 = 7.8129 and z2=4.7294z_2 = 4.7294; by (18) the predicted values of ∥rd∥2=1/(eGT(z∗)e∗)\lVert r_d\rVert^2 = 1/(e_G^{\mathsf T}(z_*)e_*) are 0.8411 and 0.7348. Table 1 compares these with direct simulations of MdM_d, n=1.2dn = 1.2d. In all 80 eigenvector comparisons the

ddrunsλ1\lambda_1λ2\lambda_2λ3\lambda_3∥rd,1∥2\lVert r_{d,1}\rVert^2∥rd,2∥2\lVert r_{d,2}\rVert^2
500207.77±0.487.77 \pm0.484.68±0.184.68 \pm0.183.05±0.083.05 \pm0.080.842±0.0140.842 \pm0.0140.734±0.0150.734 \pm0.015
1000107.60±0.297.60 \pm0.294.72±0.124.72 \pm0.123.01±0.053.01 \pm0.050.837±0.0080.837 \pm0.0080.737±0.0140.737 \pm0.014
200067.80±0.407.80 \pm0.404.80±0.144.80 \pm0.143.09±0.053.09 \pm0.050.842±0.0090.842 \pm0.0090.738±0.0080.738 \pm0.008
400047.80±0.107.80 \pm0.104.70±0.044.70 \pm0.043.09±0.023.09 \pm0.020.841±0.0050.841 \pm0.0050.733±0.0030.733 \pm0.003
prediction7.81294.7294edge 3.09760.84110.7348

Table 1. Two-class model: mean ±\pm standard deviation over independent runs of the three largest eigenvalues of MdM_d and of ∥rd∥2\lVert r_d\rVert^2 for the top two eigenvectors.

cosine of the angle between rdr_d and the predicted kernel direction e∗e_* was at least 0.98. For d≤1000d \le1000, dim⁡Ker⁡Md=d−Nactive\dim\operatorname{Ker} M_d = d - N_{\mathrm{active}} in all 30 runs, as (21) requires. The same code applied to the scalar gate of Figure 1 gives mean top eigenvalues 4.61 ±\pm 0.13, 4.68 ±\pm 0.10, 4.67 ±\pm 0.06 and 4.62 ±\pm 0.05 for d=500d = 500, 1000, 2000, 4000 (20, 10, 6, 4 runs), against 4.6653. The mean of ∥rd∥2\lVert r_d\rVert^2 lies between 0.859 and 0.862 for each dd, against 0.8609 from (18), which here coincides with the classical spiked-covariance overlap.

AI assistance disclosure

OpenAI GPT-5.6 Sol, accessed through ChatGPT and Codex, was used to develop the reduction, the numerical checks and the first draft. Anthropic Claude Opus 5.5, used through Claude Code, was used to merge and edit the note and to write the conditioning argument of § 4, including Lemma 4.4. Anthropic Claude Opus 5.5 (Claude Code) also produced Lemma 4.5, the paragraph of § 4 that replaces the unverified part of Theorem 2.9 of [3] by [6 Theorem 3.16] for the outlier assertions, and the two-class numerical check of § 12. The author reviewed and takes responsibility for every statement and proof. AI systems are not authors of this work, and the work has not been peer reviewed.

References

DEPARTMENT OF PHYSICS, ÉCOLE NORMALE SUPÉRIEURE, UNIVERSITÉ PSL, PARIS, FRANCE

Email address: samü[email protected]

References

  1. [1]Z. D. Bai and Jack W. Silverstein, No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices, Annals of Probability 26 (1998), no. 1, 316–345.
  2. [2]Jinho Baik, Gérard Ben Arous, and Sandrine Péché, Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices, Annals of Probability 33 (2005), no. 5, 1643–1697.DOI
  3. [3]Gérard Ben Arous, Reza Gheissari, Jiaoyang Huang, and Aikosh Jagannath, Local geometry of high-dimensional mixture models: Effective spectral theory and dynamical transitions, 2025, arXiv:2502.15655, version 3 (22 January 2026).arxiv.org/abs/2502.15655
  4. [4]Florent Benaych-Georges and Raj Rao Nadakuditi, The eigenvalues and eigenvectors of finite, low rank perturbations of large random matrices, Advances in Mathematics 227 (2011), no. 1, 494–521.
  5. [5]Zhou Fan, Renyuan Ma, Elliot Paquette, and Zhichao Wang, Anisotropic local law for non-separable sample covariance matrices, 2026, arXiv:2602.17960.arxiv.org/abs/2602.17960
  6. [6]Antti Knowles and Jun Yin, Anisotropic local laws for random matrices, Probability Theory and Related Fields 169 (2017), no. 1-2, 257–352, arXiv:1410.3516; statement numbers as in the arXiv version.arxiv.org/abs/1410.3516
  7. [7]Filip Kovačević, Yihan Zhang, and Marco Mondelli, Spectral estimators for multi-index models: Precise asymptotics and optimal weak recovery, Proceedings of Thirty Eighth Conference on Learning Theory, Proceedings of Machine Learning Research, vol. 291, PMLR, 2025, https://proceedings.mlr.press/v291/kovacevic25a.html, pp. 3354–3404.arxiv.org/abs/2502.01583
  8. [8]V. A. Marčenko and L. A. Pastur, Distribution of eigenvalues for some sets of random matrices, Mathematics of the USSR-Sbornik 1 (1967), no. 4, 457–483.DOI
  9. [9]Andrea Montanari and Basil Saeed, Variational formulas for the spectrum of block Wishart matrices, 2026, arXiv:2606.27774.arxiv.org/abs/2606.27774
  10. [10]Vinod Nair and Geoffrey E. Hinton, Rectified linear units improve restricted Boltzmann machines, Proceedings of the 27th International Conference on Machine Learning (Haifa, Israel), 2010.
  11. [11]Jack W. Silverstein and Sang-Il Choi, Analysis of the limiting spectral distribution of large dimensional random matrices, Journal of Multivariate Analysis 54 (1995), no. 2, 295–309.

Paper details

vlnc:W_3g3ctf6n
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

OpenAI ChatGPT GPT-5.6 Sol, OpenAI Codex GPT-5.6 Sol, Anthropic Claude Code Claude Opus 5.5

View statement
Looking for related papers…
Contents