Introduction
For a matrix M ∈ M d ( C ) M \in M_d(\mathbb{C}) M ∈ M d ( C ) with eigenvalues λ 1 ( M ) , … , λ d ( M ) \lambda_1(M),\ldots,\lambda_d(M) λ 1 ( M ) , … , λ d ( M ) , counted with algebraic multiplicity, its empirical spectral distribution is the probability measure
μ M = 1 d ∑ j = 1 d δ λ j ( M ) . \mu_M = \frac{1}{d}\sum_{j=1}^{d}\delta_{\lambda_j(M)}. μ M = d 1 j = 1 ∑ d δ λ j ( M ) .
A central problem in random matrix theory is to determine the limit of μ M N \mu_{M_N} μ M N for a sequence of random matrices M N M_N M N . For Hermitian matrices this can often be done by computing traces of powers. For nonnormal matrices, convergence of the ∗ * ∗ -moments alone does not determine the limiting eigenvalue measure. Girko’s Hermitization method [13 ] addresses this problem through the singular values of M N − z I M_N-zI M N − z I . In particular, it requires control of their logarithms near zero. The finite-variance circular law and its replacement principle provide a basic example of this method [22 ] ; we refer to [7 ] for a systematic account.
We study fixed noncommutative polynomials in independent iid matrices and their adjoints, as well as polynomials in independent Wigner matrices. For iid inputs, the entries of the polynomial need not be independent. In both models, we prove convergence of the empirical spectral distribution to the corresponding Brown measure under bounded densities and a finite ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) moment. The diagonal laws in the Wigner model need only a second moment. We prove this using a least singular value estimate for tensor pencils with arbitrary fixed coefficients. This estimate requires only second moments.
Brown measure and notation
For M ∈ M d ( C ) M \in M_d(\mathbb{C}) M ∈ M d ( C ) , set
U M ( z ) = 1 d log ∣ det ( M − z I d ) ∣ . U_M(z)=\frac{1}{d}\log\left|\det(M-zI_d)\right|. U M ( z ) = d 1 log ∣ det ( M − z I d ) ∣ .
Then μ M = ( 2 π ) − 1 Δ U M \mu_M=(2\pi)^{-1}\Delta U_M μ M = ( 2 π ) − 1 Δ U M in the sense of distributions. The corresponding measure for an operator in a tracial von Neumann algebra is the Brown measure [8 ] , defined through the Fuglede–Kadison determinant [12 ] . More precisely, let ( M , τ ) (\mathcal{M},\tau) ( M , τ ) be a tracial von Neumann algebra with faithful normalized trace. For a ∈ M a \in\mathcal{M} a ∈ M , write
U a ( z ) = τ ( log ∣ a − z ∣ ) , μ a = 1 2 π Δ U a . U_a(z)=\tau(\log|a-z|),\qquad\mu_a=\frac{1}{2\pi}\Delta U_a. U a ( z ) = τ ( log ∣ a − z ∣ ) , μ a = 2 π 1 Δ U a .
The logarithm is defined by decreasing positive regularization. Its value at a particular point may initially be − ∞ -\infty − ∞ , whereas the potential is locally integrable as a function of z z z . We use standard free circular variables c 1 , … , c n c_1,\ldots,c_n c 1 , … , c n , normalized by τ ( c ℓ c ℓ ∗ ) = 1 \tau(c_\ell c_\ell^*)=1 τ ( c ℓ c ℓ ∗ ) = 1 , and standard free semicircular variables s 1 , … , s n s_1,\ldots,s_n s 1 , … , s n , normalized by τ ( s ℓ 2 ) = 1 \tau(s_\ell^2)=1 τ ( s ℓ 2 ) = 1 . For free independence and circular and semicircular elements, see [25 , 18 ] .
All polynomials, coefficient dimensions and entry laws are fixed as N → ∞ N\to\infty N → ∞ . Superscript T denotes ordinary transpose and * denotes complex adjoint. A *-polynomial may involve the variables and their formal adjoints, with arbitrary complex coefficients. We write Tr \operatorname{Tr} Tr for the ordinary trace and tr d = d − 1 Tr \operatorname{tr}_d=d^{-1}\operatorname{Tr} tr d = d − 1 Tr for the normalized trace. For a linear map A A A , its least singular value means
s min ( A ) = inf ∥ v ∥ = 1 ∥ A v ∥ , s_{\min}(A)=\inf_{\|v\|=1}\|Av\|, s m i n ( A ) = ∥ v ∥ = 1 inf ∥ A v ∥ ,
where the infimum is over its domain. Singular values are otherwise ordered decreasingly and counted with multiplicity. Constants may depend on the fixed polynomial, coefficients and laws. We impose no coupling between the matrices at different sizes.
Independent matrices and their adjoints
We consider a centered entry law ξ \xi ξ with unit absolute variance:
E ξ = 0 , E ∣ ξ ∣ 2 = 1. (1) \mathbb{E}\xi=0,\qquad\mathbb{E}|\xi|^2=1. \tag*{(1)} E ξ = 0 , E ∣ ξ ∣ 2 = 1. (1)
The density assumption takes one of the following two forms:
(A) ξ \xi ξ is real and has a density bounded by M M M with respect to Lebesgue measure on R \mathbb{R} R ;
(B) ξ \xi ξ is complex and has a joint density bounded by M M M with respect to Lebesgue measure on C = R 2 \mathbb{C}=\mathbb{R}^2 C = R 2 .
In class (B), the real and imaginary parts may be dependent. In particular, the quantity ρ = E ξ 2 \rho=\mathbb{E}\xi^2 ρ = E ξ 2 need not vanish. Let ξ ℓ , i j \xi_{\ell,ij} ξ ℓ , ij be independent copies of ξ \xi ξ and define
X ℓ = X ℓ ( N ) = 1 N ( ξ ℓ , i j ) i , j = 1 N , 1 ≤ ℓ ≤ n . X_\ell=X_\ell^{(N)}=\frac{1}{\sqrt{N}}(\xi_{\ell,ij})_{i,j=1}^N,\qquad1\leq\ell\leq n. X ℓ = X ℓ ( N ) = N 1 ( ξ ℓ , ij ) i , j = 1 N , 1 ≤ ℓ ≤ n .
Theorem 1.1 (Brown measure convergence). *Assume either class (A) or class (B), and suppose that E ∣ ξ ∣ 2 + ε < ∞ \mathbb{E}|\xi|^{2+\varepsilon}<\infty E ∣ ξ ∣ 2 + ε < ∞ for some ε > 0 \varepsilon>0 ε > 0 . Let P P P be a fixed scalar-coefficient noncommutative -polynomial. Then
μ P ( X , X ∗ ) → P μ P ( c , c ∗ ) \mu_{P(X,X^*)}\xrightarrow{\mathbb{P}}\mu_{P(c,c^*)} μ P ( X , X ∗ ) P μ P ( c , c ∗ ) weakly, where c 1 , … , c n c_1,\ldots,c_n c 1 , … , c n is a standard free circular family. For every fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) ,
1 N log ∣ det ( P ( X , X ∗ ) − z I N ) ∣ → P τ ( log ∣ P ( c , c ∗ ) − z ∣ ) . \frac{1}{N}\log\left|\det(P(X,X^*)-zI_N)\right|\xrightarrow{\mathbb{P}}\tau\left(\log|P(c,c^*)-z|\right). N 1 log ∣ det ( P ( X , X ∗ ) − z I N ) ∣ P τ ( log ∣ P ( c , c ∗ ) − z ∣ ) .
Moreover, the logarithmic potentials converge locally in planar L 1 L^{1} L 1 in probability: for every compact K ⊂ C K \subset\mathbb{C} K ⊂ C ,
∫ K ∣ U P ( X , X ∗ ) ( z ) − U P ( c , c ∗ ) ( z ) ∣ d 2 z → P 0. \int_K \left|U_{P(\mathbf{X},\mathbf{X}^*)}(z)-U_{P(c,c^*)}(z)\right|\,\mathrm{d}^{2}z \xrightarrow{\mathbb{P}} 0. ∫ K U P ( X , X ∗ ) ( z ) − U P ( c , c ∗ ) ( z ) d 2 z P 0.
Thus the same conclusion holds for real entry laws and arbitrary complex polynomial coefficients, and for planar laws with arbitrary pseudovariance ρ \rho ρ . The least singular value theorem used to prove this result has a weaker moment assumption. To state it, fix A ℓ , B ℓ ∈ M m ( C ) A_\ell,B_\ell\in M_m(\mathbb{C}) A ℓ , B ℓ ∈ M m ( C ) and consider
T N = I m ⊗ I N + ∑ ℓ = 1 n ( A ℓ ⊗ X ℓ + B ℓ ⊗ X ℓ ∗ ) . (2) T_N=I_m\otimes I_N+\sum_{\ell=1}^{n}\left(A_\ell\otimes X_\ell+B_\ell\otimes X_\ell^*\right). \tag*{(2)} T N = I m ⊗ I N + ℓ = 1 ∑ n ( A ℓ ⊗ X ℓ + B ℓ ⊗ X ℓ ∗ ) . (2)
We shall work on the Hilbert–Schmidt event
Ω N = { max ℓ N − 1 ∥ X ℓ ∥ H S 2 ≤ 2 } , q N = P ( Ω N c ) , ω = { 1 , class (A) , 1 2 , class (B) . (3) \Omega_N=\left\{\max_\ell N^{-1}\lVert X_\ell\rVert_{\mathrm{HS}}^2\le2\right\},\qquad q_N=\mathbb{P}(\Omega_N^c),\qquad\omega=\begin{cases}1,&\text{class (A)},\\ \frac{1}{2},&\text{class (B)}.\end{cases} \tag*{(3)} Ω N = { ℓ max N − 1 ∥ X ℓ ∥ HS 2 ≤ 2 } , q N = P ( Ω N c ) , ω = { 1 , 2 1 , class (A) , class (B) . (3)
Since the entries have a second moment, the weak law of large numbers gives q N → 0 q_N\to0 q N → 0 .
Theorem 1.2 (Least singular values of tensor pencils). Under {(xref)}(1.1) and either density assumption (A) or (B), there are C , K , c > 0 C,K,c>0 C , K , c > 0 and N 0 ∈ N N_0\in\mathbb{N} N 0 ∈ N , depending on the fixed coefficients and law, such that
P { s min ( T N ) ≤ t } ≤ C N K t ω / r c o m p + C e − c N + q N , N ≥ N 0 , t > 0. (4) \mathbb{P}\{s_{\min}(T_N)\le t\}\le CN^{K}t^{\omega/r_{\mathrm{comp}}}+Ce^{-cN}+q_N,\qquad N\ge N_0,\quad t>0. \tag*{(4)} P { s m i n ( T N ) ≤ t } ≤ C N K t ω / r comp + C e − c N + q N , N ≥ N 0 , t > 0. (4) Here r c o m p ≤ m r_{\mathrm{comp}}\le m r comp ≤ m is the length of any complex composition series of C m \mathbb{C}^{m} C m for the algebra generated by I , A ℓ , B ℓ I,A_\ell,B_\ell I , A ℓ , B ℓ . For a nonzero irreducible coefficient family one may take r c o m p = 1 r_{\mathrm{comp}}=1 r comp = 1 ; one may always replace r c o m p r_{\mathrm{comp}} r comp by m m m . If E ∣ ξ ∣ 2 + ε < ∞ \mathbb{E}|\xi|^{2+\varepsilon}<\infty E ∣ ξ ∣ 2 + ε < ∞ , then q N ≤ C N − min ( ε , 2 ) q_N\le CN^{-\min(\varepsilon,2)} q N ≤ C N − m i n ( ε , 2 ) . For a real Gaussian law, or a centered complex Gaussian law with positive definite real covariance, one may take ω = 1 \omega=1 ω = 1 and replace the two additive errors by C e − c N Ce^{-cN} C e − c N .
The estimate is uniform in t t t , with constants allowed to depend on the fixed coefficient family. The denominator r c o m p r_{\mathrm{comp}} r comp arises from reconstructing an upper triangular pencil from its irreducible diagonal blocks.
Linearizing P − z I P-zI P − z I gives a tensor pencil to which Theorem 1.2 applies. Its constant matrix is invertible for z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) .
Corollary 1.3 (Least singular values of polynomial shifts). Let P P P be a fixed scalar-coefficient ∗ * ∗ -polynomial, and assume {(xref)}(1.1) and either (A) or (B). For each fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) , there are C z , a z , b z , c z > 0 C_z,a_z,b_z,c_z>0 C z , a z , b z , c z > 0 and $N_0(z) such that
P { s min ( P ( X , X ∗ ) − z I N ) ≤ t } ≤ C z N a z t b z + C z e − c z N + q N , N ≥ N 0 ( z ) , t > 0. \mathbb{P}\{s_{\min}(P(X,X^*)-zI_N)\le t\}\le C_zN^{a_z}t^{b_z}+C_ze^{-c_zN}+q_N,\qquad N\ge N_0(z),\quad t>0. P { s m i n ( P ( X , X ∗ ) − z I N ) ≤ t } ≤ C z N a z t b z + C z e − c z N + q N , N ≥ N 0 ( z ) , t > 0. In particular, for some fixed B z < ∞ B_z<\infty B z < ∞ ,
P { s min ( P ( X , X ∗ ) − z I N ) ≥ N − B z } ⟶ 1. \mathbb{P}\{s_{\min}(P(X,X^*)-zI_N)\ge N^{-B_z}\}\longrightarrow1. P { s m i n ( P ( X , X ∗ ) − z I N ) ≥ N − B z } ⟶ 1. The constants may depend on z z z . The lower cutoff for each fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) suffices for Hermitization: the logarithmic potentials are integrated over z z z , and the proof does not require uniformity in z z z .
Wigner matrices
For each 1 ≤ ℓ ≤ n 1 \le\ell\le n 1 ≤ ℓ ≤ n , let H ℓ = H ℓ ( N ) H_\ell= H_\ell^{(N)} H ℓ = H ℓ ( N ) be Hermitian with entries
( H ℓ ) i j = N − 1 / 2 ζ ℓ , i j ( i < j ) , ( H ℓ ) j i = N − 1 / 2 ζ ℓ , i j ‾ , ( H ℓ ) i i = N − 1 / 2 d ℓ , i . (H_\ell)_{ij}=N^{-1/2}\zeta_{\ell,ij}\quad(i<j),\qquad(H_\ell)_{ji}=N^{-1/2}\overline{\zeta_{\ell,ij}},\qquad(H_\ell)_{ii}=N^{-1/2}d_{\ell,i}. ( H ℓ ) ij = N − 1/2 ζ ℓ , ij ( i < j ) , ( H ℓ ) j i = N − 1/2 ζ ℓ , ij , ( H ℓ ) ii = N − 1/2 d ℓ , i .
All upper-triangle and diagonal atoms are independent, also across ℓ \ell ℓ . For each fixed ℓ \ell ℓ , the off-diagonal atoms are copies of a fixed centered law with absolute variance one. This law is either real with a bounded one-dimensional density or complex with a bounded joint planar density. Different input matrices may have different laws, and both cases may occur in the same family. The real and imaginary parts of a planar atom may be dependent. The diagonal atoms of each matrix are copies of a fixed centered real law with bounded density and finite second moment.
Theorem 1.4 (Brown measure convergence for Wigner polynomials). Assume the preceding Wigner model, and suppose that the off-diagonal laws have a finite ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) moment for some ε > 0 \varepsilon>0 ε > 0 . Let P P P be a fixed scalar-coefficient polynomial in the Hermitian variables. Then
μ P ( H 1 , … , H n ) ⟶ P μ P ( s 1 , … , s n ) \mu_{P(H_1,\ldots,H_n)} \overset{\mathbb{P}}{\longrightarrow} \mu_{P(s_1,\ldots,s_n)} μ P ( H 1 , … , H n ) ⟶ P μ P ( s 1 , … , s n ) weakly, where s 1 , … , s n s_1,\ldots,s_n s 1 , … , s n is a standard free semicircular family. For each fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) ,
1 N log ∣ det ( P ( H ) − z I N ) ∣ ⟶ P τ ( log ∣ P ( s ) − z ∣ ) . \frac{1}{N}\log\left|\det(P(H)-zI_N)\right| \overset{\mathbb{P}}{\longrightarrow} \tau\left(\log\left|P(s)-z\right|\right). N 1 log ∣ det ( P ( H ) − z I N ) ∣ ⟶ P τ ( log ∣ P ( s ) − z ∣ ) . The logarithmic potentials converge locally in planar L 1 L^1 L 1 in probability. Only a second moment is required of the diagonal laws.
For Wigner matrices, we prove the least singular value bound using their independent upper-triangle atoms. For fixed C ℓ ∈ M m ( C ) C_\ell\in M_m(\mathbb{C}) C ℓ ∈ M m ( C ) , set
T N W = I m N + ∑ ℓ = 1 n C ℓ ⊗ H ℓ , T_N^{\mathrm{W}}=I_{mN}+\sum_{\ell=1}^{n}C_\ell\otimes H_\ell, T N W = I m N + ℓ = 1 ∑ n C ℓ ⊗ H ℓ ,
and define
Ω N W = { max ℓ ≤ n N − 1 ∥ H ℓ ∥ H S 2 ≤ 2 } , q N W = P ( ( Ω N W ) c ) . (5) \Omega_N^{\mathrm{W}}=\left\{\max_{\ell\le n}N^{-1}\lVert H_\ell\rVert_{\mathrm{HS}}^2\le2\right\},\qquad q_N^{\mathrm{W}}=\mathbb{P}\left((\Omega_N^{\mathrm{W}})^c\right). \tag*{(5)} Ω N W = { ℓ ≤ n max N − 1 ∥ H ℓ ∥ HS 2 ≤ 2 } , q N W = P ( ( Ω N W ) c ) . (5)
Let ω = 1 \omega=1 ω = 1 when all off-diagonal entry laws are real, and ω = 1 / 2 \omega=1/2 ω = 1/2 when at least one has a planar density.
Theorem 1.5 (Least singular values of Wigner pencils). Under the stated density and second-moment assumptions there are C , K , c > 0 C,K,c>0 C , K , c > 0 and N 0 ∈ N N_0\in\mathbb{N} N 0 ∈ N such that
P { s min ( T N W ) ≤ t } ≤ C N K t ω / r c o m p + C e − c N + q N W , N ≥ N 0 , t > 0. \mathbb{P}\left\{s_{\min}(T_N^{\mathrm{W}})\le t\right\}\le CN^K t^{\omega/r_{\mathrm{comp}}}+Ce^{-cN}+q_N^{\mathrm{W}},\qquad N\ge N_0,\quad t>0. P { s m i n ( T N W ) ≤ t } ≤ C N K t ω / r comp + C e − c N + q N W , N ≥ N 0 , t > 0. Here r c o m p ≤ m r_{\mathrm{comp}}\le m r comp ≤ m is the length of any complex composition series of C m \mathbb{C}^m C m for the algebra generated by I , C 1 , … , C n I,C_1,\ldots,C_n I , C 1 , … , C n , and q N W → 0 q_N^{\mathrm{W}}\to0 q N W → 0 . If all atoms are Gaussian, one may take ω = 1 \omega=1 ω = 1 and replace the two additive errors by C e − c N Ce^{-cN} C e − c N .
As in Corollary 1.3, this theorem gives a polynomial least singular value cutoff for P ( H ) − z I N P(H)-zI_N P ( H ) − z I N at every fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) . The proof is given in § 6 . The single excluded shift has no effect on either weak eigenvalue convergence or local spatial L 1 L^1 L 1 convergence.
Main ideas of the proof
The proof is organized around the normalized logarithmic determinant. Write P N = P ( X , X ∗ ) P_N=P(X,X^*) P N = P ( X , X ∗ ) or P N = P ( H ) P_N=P(H) P N = P ( H ) . Its logarithmic potential is
U N ( z ) = 1 N log ∣ det ( P N − z I N ) ∣ = 1 N ∑ j = 1 N log s j ( P N − z I N ) , μ P N = 1 2 π Δ U N . U_N(z)=\frac{1}{N}\log\left|\det(P_N-zI_N)\right|=\frac{1}{N}\sum_{j=1}^{N}\log s_j(P_N-zI_N),\qquad\mu_{P_N}=\frac{1}{2\pi}\Delta U_N. U N ( z ) = N 1 log ∣ det ( P N − z I N ) ∣ = N 1 j = 1 ∑ N log s j ( P N − z I N ) , μ P N = 2 π 1 Δ U N .
Thus the limiting eigenvalue measure can be identified through local L 1 L^1 L 1 convergence of U N U_N U N to the logarithmic potential U U U of the free polynomial. To justify the logarithmic integral, weak singular-value convergence must be supplemented by estimates for very small and very large singular values. Figure 1 shows how the estimates address these tails.
Figure 1. From singular-value estimates to logarithmic potentials and Brown measures. Trace observations control inverse rows; Gaussian/free estimates and atom comparison; Second moments and rank truncation; Logarithmic derivatives and the LSV bound; Intermediate small singular-value count; Weak convergence and upper log tail; Normalized logarithmic determinant, U N ( z ) → P U ( z ) U_N(z)\xrightarrow{\mathbb{P}}U(z) U N ( z ) P U ( z ) for each fixed z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) ; Local L 1 L^1 L 1 convergence of logarithmic potentials using spatial uniform integrability; Brown measure convergence via the distributional Laplacian.
We follow the polynomial linearization and irreducible reduction of [15 Sections 2 and 3.1] . Its Gaussian and non-Gaussian results treat polynomials without adjoints. On each nontrivial factor, we must show that traces against the coefficients control the size of an inverse row. § 2 constructs the required trace observations; § 3 adapts the dimension-surplus covering argument of [15 ] to the shared real atom coordinates. The inverse equations then give a bound for every inverse row. Summing these bounds and applying Jacobi’s formula yields, for F = det T N F=\det T_N F = det T N ,
∥ T N − 1 ∥ H S ≤ N C ∥ ∇ F ∥ 2 ∣ F ∣ . \left\|T_N^{-1}\right\|_{\mathrm{HS}}\le N^C\frac{\|\nabla F\|_2}{|F|}. T N − 1 HS ≤ N C ∣ F ∣ ∥∇ F ∥ 2 .
on Ω N \Omega_N Ω N outside an event of probability C e − c N Ce^{-cN} C e − c N . Figure 2 displays the calculation behind this bound.
Figure 2. Logarithmic derivatives convert inverse traces into determinant anti-concentration. Only the affected spatial indices are shown. Here ι \iota ι ranges over the original real coordinates, and E N ⊂ Ω N \mathcal{E}_N \subset\Omega_N E N ⊂ Ω N is the frame event with exponent β \beta β for a nonzero irreducible coefficient family. The zero set of F F F is omitted. For i = j i=j i = j , the two perturbed blocks merge into A + B A+B A + B .
This is the main departure from the least singular value argument in [15 Sections 3.4–3.7 and 6.1] . There one conditions on all but one block column, estimates the determinant of its fixed-size projection, and applies the block negative second moment identity. Adjoint and Hermitian couplings make different block columns share atoms. Here the trace observations give derivatives of the full determinant in the original real coordinates. Relative polynomial anti-concentration then gives the least singular value bound in § 4 . The triangular reconstruction from [15 Section 3.1] combines the factors in § 5 ; § 6 treats the coupling at the two endpoints of a Wigner edge.
To prove convergence of U N U_N U N , we combine the least singular value cutoff with a count at intermediate scales. Gaussian/free comparison and rank truncation give this count, and together the bounds control the lower logarithmic tail. Second moments control the upper tail. § 7 combines these estimates with weak singular-value convergence to obtain pointwise convergence of U N U_N U N . A uniform expected spatial L 2 L^2 L 2 bound gives uniform integrability and local L 1 L^1 L 1 convergence; the Laplacian then identifies the Brown measure. Thus the pencil determinant is used through its logarithmic derivatives to control the inverse, and the polynomial determinant through its normalized logarithm to identify the eigenvalue measure.
Relation to previous work
The free limits appearing here come from Voiculescu’s asymptotic freeness theorem [24 ] . For independent GUE matrices, Haagerup and Thorbjørnsen [14 ] also proved convergence of polynomial norms. Male [17 ] extended this framework to the presence of independent matrices with a suitable strong limit, and Anderson [2 ] proved polynomial norm convergence for independent Wigner matrices under fourth-moment hypotheses. Belinschi, Bercovici and Capitaine [4 ] studied outliers of self-adjoint polynomial deformations under spectral assumptions on the input matrices. These results give norm convergence and spectral exclusion in their respective settings. Convergence of the eigenvalue measure for a nonnormal polynomial additionally requires control of the singular logarithm. Operator-valued subordination gives analytic descriptions of free polynomial distributions [5 ] . Linearizations and realizations are developed in [16 ] , while Hermitian reduction provides a method for studying Brown measures of nonnormal free polynomials [6 ] . Cook, Guionnet and Husson [10 ] proved Brown measure convergence for quadratic polynomials in independent complex Ginibre matrices. The author’s June 2026 paper [15 ] established this convergence for polynomials of arbitrary degree without adjoints in both Ginibre matrices and non-Gaussian iid matrices. The non-Gaussian convergence holds under bounded planar density and finite moments of all orders; see [15 Theorem 1.4 and Section 6] . Its coefficient-module method also gives least singular value estimates for linearized tensor pencils in both settings, with the non-Gaussian estimate proved in [15 Theorem 6.3] .
Collins, Kitagawa, Miyagawa, Yuan and Zhong [9 ] subsequently proved weak Brown measure convergence for arbitrary GUE polynomials, equivalently for arbitrary *-polynomials in independent complex Ginibre matrices. Their least singular value theorem for Ginibre pencils with structured means extends the Gaussian case of the centered tensor-pencil estimates in [15 Theorems 1.5 and 6.3] , following its coefficient-space and restricted-adjoint argument; see [9 Sections 1.3, 3.3 and 4] . They combine this extension with an auxiliary noisy ensemble, a second-moment criterion for Brown measures, Gaussian rotation and convexity. They also established convergence of the spectral set and spectral radius, as well as a free support theorem.
The polynomial convergence results in [9 Theorems 1.1 and 1.4] cover Gaussian inputs only: independent GUE matrices, or equivalently independent complex Ginibre matrices with adjoints. General Wigner matrices are explicitly left untreated in [9 Section 1.1] . This restriction enters the proof of [9 Theorem 3.3] : real rotations of two independent GUE families preserve their joint law, making the relevant subharmonic function independent of the real part of the rotation parameter and convex in its imaginary part. For non-Gaussian entry laws, the same rotation can change the joint law. The bounded densities and moment assumptions considered here do not imply this invariance, so the rotation step does not yield the non-Gaussian convergence theorem proved below.
The present proof develops the coefficient-module reduction and restricted-adjoint argument of [15 ] . The doubled coefficient algebra and its adjoint space give inverse-row trace observations that control the full determinant gradient. This replaces conditional block exposure and treats adjoint and Hermitian couplings. The least singular value proof uses neither Gaussian rotation nor an auxiliary Gaussian perturbation, and applies to real entry laws under the stated moment assumptions.
Our probability estimates are also related to the scalar invertibility theory. Littlewood–Offord and small-ball methods were developed for independent-entry matrices in [19 , 20 ] , and decoupling methods for symmetric matrices in [23 ] . Under bounded densities, Farrell and Vershynin [11 ] obtained polynomial invertibility bounds for symmetric matrices without a moment assumption, uniformly over deterministic symmetric additions. Here an atom can enter several coefficient blocks, and the resulting tensor pencil can be nonnormal. The coefficient algebra and the trace estimate for its inverse blocks allow us to treat this dependence.
The Hermitian pencils are also related to the Kronecker random matrices studied in [1 ] . The spectral comparison uses the GUE resolvent estimate of Anderson, Guionnet and Zeitouni [3 Lemma 5.5.4] and the non-atomicity and algebraicity results of Shlyakhtenko and Skoufranis [21 Theorem 1.1 and Corollary 1.2] . The former estimates the Gaussian resolvents; the latter gives the small-singular-value bound for the free limit.
Real coordinates
We treat classes (A) and (B) in a common notation. Set d = 1 d=1 d = 1 in class (A) and d = 2 d=2 d = 2 in class (B), and write
u ℓ , i j = ( u ℓ , i j , h ) h = 1 d , ξ ℓ , i j = ∑ h = 1 d κ h u ℓ , i j , h , κ 1 = 1 , κ 2 = i . (6) u_{\ell,ij}=\left(u_{\ell,ij,h}\right)_{h=1}^{d},\qquad\xi_{\ell,ij}=\sum_{h=1}^{d}\kappa_hu_{\ell,ij,h},\qquad\kappa_1=1,\quad\kappa_2=i. \tag*{(6)} u ℓ , ij = ( u ℓ , ij , h ) h = 1 d , ξ ℓ , ij = h = 1 ∑ d κ h u ℓ , ij , h , κ 1 = 1 , κ 2 = i . (6)
Here u = ξ u=\xi u = ξ in class (A) and u = ( ℜ ξ , ℑ ξ ) u=(\Re\xi,\Im\xi) u = ( ℜ ξ , ℑ ξ ) in class (B). These independent blocks have bounded joint densities and $\mathbb{E}|u|^2=1.
Define the real matrices Y ℓ , h = N − 1 / 2 ( u ℓ , i j , h ) i j Y_{\ell,h}=N^{-1/2}(u_{\ell,ij,h})_{ij} Y ℓ , h = N − 1/2 ( u ℓ , ij , h ) ij and the expanded coefficients
A ℓ , h = κ h A ℓ , B ℓ , h = κ h ‾ B ℓ . (7) A_{\ell,h}=\kappa_hA_\ell,\qquad B_{\ell,h}=\overline{\kappa_h}B_\ell. \tag*{(7)} A ℓ , h = κ h A ℓ , B ℓ , h = κ h B ℓ . (7)
Then (2) becomes
T N = I + ∑ ℓ , h ( A ℓ , h ⊗ Y ℓ , h + B ℓ , h ⊗ Y ℓ , h T ) . (8) T_N=I+\sum_{\ell,h}\left(A_{\ell,h}\otimes Y_{\ell,h}+B_{\ell,h}\otimes Y_{\ell,h}^{\mathsf T}\right). \tag*{(8)} T N = I + ℓ , h ∑ ( A ℓ , h ⊗ Y ℓ , h + B ℓ , h ⊗ Y ℓ , h T ) . (8)
In particular, the imaginary coordinate contributes i A ℓ ⊗ Y ℓ , 2 − i B ℓ ⊗ Y ℓ , 2 T iA_\ell\otimes Y_{\ell,2}-iB_\ell\otimes Y_{\ell,2}^{\mathsf T} i A ℓ ⊗ Y ℓ , 2 − i B ℓ ⊗ Y ℓ , 2 T . The expanded coefficients generate the same complex algebra as A ℓ , B ℓ A_\ell,B_\ell A ℓ , B ℓ : the h = 1 h=1 h = 1 coefficients already contain this family, and the others are scalar multiples.
The real covariance of a planar atom is
Σ ρ = 1 2 ( 1 + ℜ ρ ℑ ρ ℑ ρ 1 − ℜ ρ ) . (9) \Sigma_\rho=\frac{1}{2}\begin{pmatrix}1+\Re\rho&\Im\rho\\ \Im\rho&1-\Re\rho\end{pmatrix}. \tag*{(9)} Σ ρ = 2 1 ( 1 + ℜ ρ ℑ ρ ℑ ρ 1 − ℜ ρ ) . (9)
The bound ∣ ρ ∣ ≤ 1 |\rho|\le1 ∣ ρ ∣ ≤ 1 follows from (1) . In class (B), Σ ρ \Sigma_\rho Σ ρ is positive definite and hence ∣ ρ ∣ < 1 |\rho|<1 ∣ ρ ∣ < 1 . Indeed, a zero-variance real linear form would force the law to be supported on a line, contradicting its planar density. Class (A) corresponds to ρ = 1 \rho=1 ρ = 1 and a one-dimensional real atom.
In class (A), X ℓ ∗ = X ℓ T X_\ell^*=X_\ell^{\mathsf T} X ℓ ∗ = X ℓ T . For a complex matrix space viewed as a real Hilbert space we use
⟨ U , V ⟩ R = ℜ Tr ( U ∗ V ) . \langle U,V\rangle_{\mathbb{R}}=\Re\operatorname{Tr}(U^*V). ⟨ U , V ⟩ R = ℜ Tr ( U ∗ V ) .
Spatial blocks are taken after the fixed unitary tensor flip that places the spatial indices first. In particular, the spatial blocks of an m N × m N mN\times mN m N × m N tensor pencil are m × m m\times m m × m matrices.
Organization
Section 2 identifies the trace observations for linearized pencils. Section 3 proves the probability estimate that allows them to control inverse rows. Section 4 passes from these rows to the determinant gradient and the least singular value. Sections 5 and 6 extend the bound to arbitrary coefficients and Wigner inputs. Section 7 controls the logarithmic tails and proves Brown measure convergence.
AI assistance. The author used GPT-6 Astra and GPT-6.1 Sol to assist with algebraic questions and some technical aspects of the ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) -moment arguments.
Linearization and trace observations
We first linearize the polynomial, following [15 Section 2] . The new task is to identify traces that control inverse rows when an atom enters two transposed positions. A doubled coefficient algebra expresses these paired traces. The resulting trace kernel satisfies the invariant-subspace condition required in § 3 .
Linearization of polynomial shifts
We follow the reduction from polynomial shifts to normalized tensor pencils in [15 Section 2] . The word-chain construction below also gives polynomial elimination matrices; see [2 , 5 , 16 ] for the general theory. A matrix-valued ∗ \ast ∗ -polynomial has a fixed number of blocks, each of which is a fixed scalar-coefficient noncommutative ∗ \ast ∗ -polynomial in Z , Z ∗ Z, Z^\ast Z , Z ∗ . Evaluation at N × N N \times N N × N matrices therefore gives N × N N \times N N × N blocks. The block number, degrees and coefficients do not depend on N N N .
Lemma 2.1 (Explicit word-chain linearization). For every nonconstant P P P with c 0 = P ( 0 ) c_0=P(0) c 0 = P ( 0 ) there is a fixed-size affine ∗ \ast ∗ -pencil
L z ( Z ) = K z ⊗ I + ∑ ℓ ( U ℓ ⊗ Z ℓ + V ℓ ⊗ Z ℓ ∗ ) \mathcal{L}_z(Z)=K_z\otimes I+\sum_{\ell}(U_\ell\otimes Z_\ell+V_\ell\otimes Z_\ell^\ast) L z ( Z ) = K z ⊗ I + ℓ ∑ ( U ℓ ⊗ Z ℓ + V ℓ ⊗ Z ℓ ∗ ) such that K z K_z K z is invertible for z ≠ c 0 z\ne c_0 z = c 0 ; the upper-left inverse block of L z \mathcal{L}_z L z is ( P ( Z , Z ∗ ) − z I ) − 1 (P(Z,Z^\ast)-zI)^{-1} ( P ( Z , Z ∗ ) − z I ) − 1 whenever it exists; and there are invertible matrix-valued ∗ \ast ∗ -polynomials V z , W z V_z,W_z V z , W z with
V z L z W z = diag ( P ( Z , Z ∗ ) − z I , I ) . V_z\mathcal{L}_zW_z=\operatorname{diag}(P(Z,Z^\ast)-zI,I). V z L z W z = diag ( P ( Z , Z ∗ ) − z I , I ) . Both elimination matrices and their inverses are fixed matrix-valued ∗ \ast ∗ -polynomials, uniformly bounded when input norms are bounded. These are algebraic identities for every tuple of complex matrices Z Z Z and thus apply to both input classes with Z = X Z=X Z = X .
Proof. Write P = c 0 + ∑ w a w w P=c_0+\sum_w a_w w P = c 0 + ∑ w a w w , where the sum contains only nonempty words with nonzero coefficients. For w = Z i 1 ε 1 ⋯ Z i d w ε d w w=Z_{i_1}^{\varepsilon_1}\cdots Z_{i_{d_w}}^{\varepsilon_{d_w}} w = Z i 1 ε 1 ⋯ Z i d w ε d w , with each ε j \varepsilon_j ε j either 1 1 1 or ∗ \ast ∗ , let S w S_w S w have size d w + 1 d_w+1 d w + 1 and first superdiagonal
( S w ) j , j + 1 = Z i j ε j , 1 ≤ j ≤ d w , (S_w)_{j,j+1}=Z_{i_j}^{\varepsilon_j},\qquad1\le j\le d_w, ( S w ) j , j + 1 = Z i j ε j , 1 ≤ j ≤ d w ,
with all other blocks zero. Nilpotence gives
( I − S w ) − 1 = ∑ k = 0 d w S w k , ( ( I − S w ) − 1 ) 1 , d w + 1 = w . (I-S_w)^{-1}=\sum_{k=0}^{d_w}S_w^k,\qquad((I-S_w)^{-1})_{1,d_w+1}=w. ( I − S w ) − 1 = k = 0 ∑ d w S w k , (( I − S w ) − 1 ) 1 , d w + 1 = w .
Set Q = ⨁ w ( I − S w ) Q=\bigoplus_w(I-S_w) Q = ⨁ w ( I − S w ) . Let u u u be the block row with − a w I -a_wI − a w I in the first coordinate of each word chain and zero elsewhere, and let v v v be the block column with I I I in the last coordinate of each chain. Then
u Q − 1 v = − ∑ w a w w = c 0 − P . uQ^{-1}v=-\sum_w a_ww=c_0-P. u Q − 1 v = − w ∑ a w w = c 0 − P .
In the following matrices scalar constants are tensored with the identity of the input dimension. Define
L z = ( c 0 − z u v Q ) , V z = ( I 0 0 Q − 1 ) ( I − u Q − 1 0 I ) , W z = ( I 0 − Q − 1 v I ) . \mathcal{L}_z=\begin{pmatrix}c_0-z&u\\v&Q\end{pmatrix},\qquad
V_z=\begin{pmatrix}I&0\\0&Q^{-1}\end{pmatrix}
\begin{pmatrix}I&-uQ^{-1}\\0&I\end{pmatrix},\qquad
W_z=\begin{pmatrix}I&0\\-Q^{-1}v&I\end{pmatrix}. L z = ( c 0 − z v u Q ) , V z = ( I 0 0 Q − 1 ) ( I 0 − u Q − 1 I ) , W z = ( I − Q − 1 v 0 I ) .
Block multiplication, using u Q − 1 v = c 0 − P uQ^{-1}v=c_0-P u Q − 1 v = c 0 − P , gives
L z W z = ( P − z I u 0 Q ) , V z L z W z = ( P − z I 0 0 I ) . \mathcal{L}_zW_z=\begin{pmatrix}P-zI&u\\0&Q\end{pmatrix},\qquad
V_z\mathcal{L}_zW_z=\begin{pmatrix}P-zI&0\\0&I\end{pmatrix}. L z W z = ( P − z I 0 u Q ) , V z L z W z = ( P − z I 0 0 I ) .
Whenever P − z I P-zI P − z I is invertible, this identity gives
L z − 1 = W z ( ( P − z I ) − 1 0 0 I ) V z . \mathcal{L}_z^{-1}=W_z\begin{pmatrix}(P-zI)^{-1}&0\\0&I\end{pmatrix}V_z. L z − 1 = W z ( ( P − z I ) − 1 0 0 I ) V z .
The first block row of W z W_z W z is ( I , 0 ) (I,0) ( I , 0 ) and the first block column of V z V_z V z is ( I , 0 ) T (I,0)^{\mathsf T} ( I , 0 ) T , which gives the upper-left block of the inverse. The finite geometric series for Q − 1 Q^{-1} Q − 1 shows that V z V_z V z and W z W_z W z are matrix-valued ∗ \ast ∗ -polynomials. Their unitriangular factors invert by changing the off-diagonal sign, while diag ( I , Q − 1 ) − 1 = \operatorname{diag}(I,Q^{-1})^{-1}= diag ( I , Q − 1 ) − 1 = diag( I , Q ) (I,Q) ( I , Q ) . Their inverses are also matrix-valued *-polynomials, so all four matrices are uniformly bounded when the input norms are bounded. At zero substitution Q = I Q=I Q = I and u v = 0 uv=0 uv = 0 , since the first and last coordinates of each nonempty word chain are distinct. The scalar Schur complement then gives det K z = c 0 − z \det K_z=c_0-z det K z = c 0 − z , so K z K_z K z is invertible exactly when z ≠ c 0 z\ne c_0 z = c 0 . □
Proof of Corollary 1.3. For constant P = c 0 P=c_0 P = c 0 , the least singular value equals the fixed positive number ∣ c 0 − z ∣ \lvert c_0-z\rvert ∣ c 0 − z ∣ , so the assertion is immediate after choosing the constant in the bound. Assume henceforth that P P P is nonconstant. Normalize by T z , N = ( K z − 1 ⊗ I ) L z ( X ) T_{z,N}=(K_z^{-1}\otimes I)L_z(X) T z , N = ( K z − 1 ⊗ I ) L z ( X ) . This is a pencil of the form in Theorem 1.2 , with fixed complex coefficients depending on z z z . Since
∥ ( P − z I ) − 1 ∥ ≤ ∥ L z − 1 ∥ ≤ ∥ T z , N − 1 ∥ ∥ K z − 1 ∥ , \lVert(P-zI)^{-1}\rVert\leq\lVert L_z^{-1}\rVert\leq\lVert T_{z,N}^{-1}\rVert\lVert K_z^{-1}\rVert, ∥( P − z I ) − 1 ∥ ≤ ∥ L z − 1 ∥ ≤ ∥ T z , N − 1 ∥ ∥ K z − 1 ∥ ,
the desired bound follows. □
Remark 2.2 (A singular shift in the real skew-symmetric model). The exceptional shift cannot be removed uniformly over all *-polynomials in the real model. For P ( x , x ∗ ) = x − x ∗ P(x,x^*)=x-x^* P ( x , x ∗ ) = x − x ∗ , one has P ( 0 ) = 0 P(0)=0 P ( 0 ) = 0 , and P ( X , X T ) P(X,X^{\mathsf T}) P ( X , X T ) is real skew-symmetric. In every odd dimension its determinant is zero. Excluding this scalar shift does not affect the almost-everywhere argument in the plane for Brown measure convergence.
The doubled coefficient algebra
Assume that ( A ℓ , B ℓ ) (A_\ell,B_\ell) ( A ℓ , B ℓ ) is nonzero and complex irreducible. The determinant derivatives pair traces at spatial indices ( i , j ) (i,j) ( i , j ) and ( j , i ) (j,i) ( j , i ) . We adjoin a transpose copy of the pencil to express these pairs as traces in one coefficient algebra. Its adjoint space will be used in Theorem 3.1.
By (7) , the expanded tuple generates the same algebra as the original tuple, so it is still nonzero and irreducible. For every ( ℓ , h ) (\ell,h) ( ℓ , h ) define
C ℓ , h , 1 = ( A ℓ , h 0 0 B ℓ , h T ) , C ℓ , h , 2 = ( B ℓ , h 0 0 A ℓ , h T ) , B = alg C ( I , C ℓ , h , 1 , C ℓ , h , 2 ) . (10) C_{\ell,h,1}=
\begin{pmatrix}
A_{\ell,h} & 0\\
0 & B_{\ell,h}^{\mathsf T}
\end{pmatrix},
\qquad
C_{\ell,h,2}=
\begin{pmatrix}
B_{\ell,h} & 0\\
0 & A_{\ell,h}^{\mathsf T}
\end{pmatrix},
\qquad
\mathcal{B}=\operatorname{alg}_{\mathbb{C}}(I,C_{\ell,h,1},C_{\ell,h,2}). \tag*{(10)} C ℓ , h , 1 = ( A ℓ , h 0 0 B ℓ , h T ) , C ℓ , h , 2 = ( B ℓ , h 0 0 A ℓ , h T ) , B = alg C ( I , C ℓ , h , 1 , C ℓ , h , 2 ) . (10)
The associated pencil, after the fixed coordinate ordering, is
T ^ N = I + ∑ ℓ , h ( C ℓ , h , 1 ⊗ Y ℓ , h + C ℓ , h , 2 ⊗ Y ℓ , h T ) ≃ T N ⊕ T N T . (11) \widehat{T}_N=I+\sum_{\ell,h}\left(C_{\ell,h,1}\otimes Y_{\ell,h}+C_{\ell,h,2}\otimes Y_{\ell,h}^{\mathsf T}\right)\simeq T_N\oplus T_N^{\mathsf T}. \tag*{(11)} T N = I + ℓ , h ∑ ( C ℓ , h , 1 ⊗ Y ℓ , h + C ℓ , h , 2 ⊗ Y ℓ , h T ) ≃ T N ⊕ T N T . (11)
Indeed, its lower diagonal block is
I + ∑ ℓ , h ( B ℓ , h T ⊗ Y ℓ , h + A ℓ , h T ⊗ Y ℓ , h T ) = T N T . I+\sum_{\ell,h}\left(B_{\ell,h}^{\mathsf T}\otimes Y_{\ell,h}+A_{\ell,h}^{\mathsf T}\otimes Y_{\ell,h}^{\mathsf T}\right)=T_N^{\mathsf T}. I + ℓ , h ∑ ( B ℓ , h T ⊗ Y ℓ , h + A ℓ , h T ⊗ Y ℓ , h T ) = T N T .
The transpose in (11) is therefore the ordinary transpose of T N T_N T N , with X ∗ X^* X ∗ expressed in the expanded real coordinates. The algebra B \mathcal{B} B is a complex subalgebra of M m ( C ) ⊕ M m ( C ) M_m(\mathbb{C})\oplus M_m(\mathbb{C}) M m ( C ) ⊕ M m ( C ) and has complex dimension at most 2 m 2 2m^2 2 m 2 .
Lemma 2.3 (Trace pairing and common kernel). The bilinear trace form ( b , c ) ↦ Tr ( b c ) (b,c)\mapsto\operatorname{Tr}(bc) ( b , c ) ↦ Tr ( b c ) on B \mathcal{B} B is nondegenerate. Moreover ⋂ ℓ , h , t ker C ℓ , h , t = { 0 } \bigcap_{\ell,h,t}\ker C_{\ell,h,t}=\{0\} ⋂ ℓ , h , t ker C ℓ , h , t = { 0 } .
Proof. The first coordinate representation of B \mathcal{B} B on C m \mathbb{C}^m C m is irreducible by hypothesis. The second is also irreducible: if a subspace is invariant under all B ℓ , h T , A ℓ , h T B_{\ell,h}^{\mathsf T},A_{\ell,h}^{\mathsf T} B ℓ , h T , A ℓ , h T , its annihilator for the bilinear pairing u T v u^{\mathsf T}v u T v is invariant under all B ℓ , h , A ℓ , h B_{\ell,h},A_{\ell,h} B ℓ , h , A ℓ , h . Thus the natural faithful B \mathcal{B} B -module is a direct sum of two irreducible modules. The radical annihilates each simple module, so faithfulness forces the radical to vanish. Hence B \mathcal{B} B is semisimple. In a faithful representation of a finite-dimensional complex semisimple algebra ⨁ j M d j ( C ) \bigoplus_j M_{d_j}(\mathbb{C}) ⨁ j M d j ( C ) , the ambient trace is ∑ j r j Tr d j \sum_j r_j\operatorname{Tr}_{d_j} ∑ j r j Tr d j with positive integer multiplicities r j r_j r j . If an element has a nonzero component in M d j ( C ) M_{d_j}(\mathbb{C}) M d j ( C ) , pairing that component with a suitable matrix unit gives a nonzero trace. Since r j r_j r j is positive, the trace form on B \mathcal{B} B is nondegenerate.
The common kernel of a nonzero irreducible family is zero: it is an invariant subspace and cannot be the entire space. Apply this to both coordinate families to obtain the last assertion.
Inverse blocks and trace observations
The inverse blocks belong to the doubled algebra. Its traces pair the two spatial positions changed by a single atom, as required for determinant differentiation.
Lemma 2.4 (Inverse blocks and the paired observations). If T N T_N T N is invertible and R = T N − 1 R=T_N^{-1} R = T N − 1 , every block R ^ i j \widehat{R}_{ij} R ij of R ^ = T ^ N − 1 \widehat{R}=\widehat{T}_N^{-1} R = T N − 1 belongs to B \mathcal{B} B , and
R ^ i j = diag ( R i j , R j i T ) , (12) \widehat{R}_{ij}=\operatorname{diag}(R_{ij},R_{ji}^{\mathsf T}), \tag*{(12)} R ij = diag ( R ij , R j i T ) , (12)
Tr ( R ^ j i C ℓ , h , 1 ) = Tr ( R j i A ℓ , h ) + Tr ( R i j B ℓ , h ) , (13) \operatorname{Tr}(\widehat{R}_{ji}C_{\ell,h,1})=\operatorname{Tr}(R_{ji}A_{\ell,h})+\operatorname{Tr}(R_{ij}B_{\ell,h}), \tag*{(13)} Tr ( R j i C ℓ , h , 1 ) = Tr ( R j i A ℓ , h ) + Tr ( R ij B ℓ , h ) , (13)
Tr ( R ^ j i C ℓ , h , 2 ) = Tr ( R j i B ℓ , h ) + Tr ( R i j A ℓ , h ) . (14) \operatorname{Tr}(\widehat{R}_{ji}C_{\ell,h,2})=\operatorname{Tr}(R_{ji}B_{\ell,h})+\operatorname{Tr}(R_{ij}A_{\ell,h}). \tag*{(14)} Tr ( R j i C ℓ , h , 2 ) = Tr ( R j i B ℓ , h ) + Tr ( R ij A ℓ , h ) . (14)
:::
Proof. The block formula follows from (11) . Cayley–Hamilton writes T ^ N − 1 \widehat{T}_N^{-1} T N − 1 as a polynomial in T ^ N \widehat{T}_N T N ; its blocks therefore belong to the unital algebra B \mathcal{B} B . Multiplying the diagonal blocks and using cyclicity and invariance of trace under transpose gives both trace identities.
The trace kernel and invariant subspaces
To control an inverse row by these traces, we must exclude profiles concentrated in their common kernel. Let V = B ∗ : = { b ∗ : b ∈ B } V=\mathcal{B}^*:=\{b^*:b\in\mathcal{B}\} V = B ∗ := { b ∗ : b ∈ B } , with the real Hilbert–Schmidt inner product, and define
O ( Φ ) = ( Tr ( Φ ∗ C ℓ , h , t ) ) ℓ , h , t , D = ker O ⊂ V . (15) O(\Phi)=\left(\operatorname{Tr}(\Phi^*C_{\ell,h,t})\right)_{\ell,h,t},\qquad\mathcal{D}=\ker O\subset V. \tag*{(15)} O ( Φ ) = ( Tr ( Φ ∗ C ℓ , h , t ) ) ℓ , h , t , D = ker O ⊂ V . (15)
This map is real linear and complex conjugate linear, so its kernel is a complex subspace. Adjoints of maps on V V V below are real Hilbert adjoints, including orthogonal compression to V V V when the map is defined on the ambient matrix space; no closure of B \mathcal{B} B under matrix adjoint is assumed. Also, for Φ = b ∗ ∈ V \Phi=b^*\in V Φ = b ∗ ∈ V ,
C ℓ , h , t ∗ Φ = ( b C ℓ , h , t ) ∗ ∈ V . C_{\ell,h,t}^*\Phi=(bC_{\ell,h,t})^*\in V. C ℓ , h , t ∗ Φ = ( b C ℓ , h , t ) ∗ ∈ V .
Thus left multiplication acts on V V V . The following lemma verifies the structural hypothesis of Theorem 3.1.
Lemma 2.5 (Trace observability on the adjoint algebra). No nonzero real E ⊂ D E\subset\mathcal{D} E ⊂ D is invariant under all left multiplications by C ℓ , h , t ∗ C_{\ell,h,t}^* C ℓ , h , t ∗ .
Proof. Suppose such a space E E E exists. Since D \mathcal{D} D is complex, E C = E + i E E_{\mathbb{C}}=E+iE E C = E + i E is a nonzero complex subspace of D \mathcal{D} D . Invariance under the generators implies invariance under the unital algebra B ∗ \mathcal{B}^* B ∗ acting on the left. Taking adjoints reverses multiplication, so
J = { Φ ∗ : Φ ∈ E C } ⊂ B , W b ∈ J ( W ∈ J , b ∈ B ) . J=\{\Phi^*:\Phi\in E_{\mathbb{C}}\}\subset\mathcal{B},\qquad Wb\in J\quad(W\in J,\ b\in\mathcal{B}). J = { Φ ∗ : Φ ∈ E C } ⊂ B , W b ∈ J ( W ∈ J , b ∈ B ) .
Thus J J J is a nonzero right ideal. Since every W b Wb W b is the adjoint of an element of D \mathcal{D} D , for each generator we have
Tr ( C ℓ , h , t W b ) = Tr ( W b C ℓ , h , t ) = 0 for every b ∈ B . \operatorname{Tr}(C_{\ell,h,t}Wb)=\operatorname{Tr}(WbC_{\ell,h,t})=0\qquad\text{for every }b\in\mathcal{B}. Tr ( C ℓ , h , t W b ) = Tr ( W b C ℓ , h , t ) = 0 for every b ∈ B .
The element C ℓ , h , t W C_{\ell,h,t}W C ℓ , h , t W belongs to B \mathcal{B} B . Nondegeneracy of the trace pairing therefore gives C ℓ , h , t W = 0 C_{\ell,h,t}W=0 C ℓ , h , t W = 0 . Every column of W W W lies in the common kernel from Lemma 2.3 , so W = 0 W=0 W = 0 . This contradicts J ≠ 0 J\ne0 J = 0 . □
A probability estimate for the trace kernel
The dimension surplus and multiscale covering argument come from [15 Theorem 3.3, Section 4 and Proposition 6.2] . Here we adapt the probability estimates to whole real atom blocks: coordinates within a block may be dependent, and the same atom occurs in both row and column terms of the inverse equations.
We prove a polynomial least singular value bound for those equations on the trace kernel. Taking the adjoint of an inverse row gives a vector of coefficient blocks, called a profile, which solves every equation except one. The preceding algebraic argument supplies the structural condition that prevents such a profile from concentrating in the trace kernel. The next section uses this bound to control the full inverse through its traces.
For a fixed profile, anchor exposure leaves independent whole blocks in different outputs. The dimension surplus allows summation over the profile net. Spectral scales handle small singular values of the coefficient map; projection handles small eigenvalues of the profile Gram operator.
Let V V V be a real Hilbert space of fixed dimension p ≥ 1 p\ge1 p ≥ 1 , let D ⊂ V \mathcal{D}\subset V D ⊂ V be a fixed subspace, and let L ℓ , h , M ℓ , h : V → V L_{\ell,h},M_{\ell,h}:V\to V L ℓ , h , M ℓ , h : V → V be fixed real linear maps. Write P E P_E P E for the orthogonal projection onto E ⊂ V E\subset V E ⊂ V and Gr ( d , D ) \operatorname{Gr}(d,\mathcal{D}) Gr ( d , D ) for the Grassmannian of d d d -dimensional real subspaces of D \mathcal{D} D . We use the independent blocks u ℓ , i j ∈ R d u_{\ell,ij}\in\mathbb{R}^d u ℓ , ij ∈ R d of (6) , whose joint densities are bounded and whose second moments are finite. The coordinates of a single block may be dependent. Our structural assumption is
no nonzero real E ⊂ D is invariant under every L ℓ , h , M ℓ , h . (16) \text{no nonzero real }E\subset\mathcal{D}\text{ is invariant under every }L_{\ell,h},M_{\ell,h}. \tag*{(16)} no nonzero real E ⊂ D is invariant under every L ℓ , h , M ℓ , h . (16)
For a deleted output a ∈ [ N ] a\in[N] a ∈ [ N ] , define
( D a v ) k = v k + 1 N ∑ i , ℓ , h ( u ℓ , i k , h L ℓ , h v i + u ℓ , k i , h M ℓ , h v i ) , k ≠ a , v ∈ D N . (17) (\mathcal{D}_a v)_k=v_k+\frac{1}{\sqrt{N}}\sum_{i,\ell,h}\left(u_{\ell,ik,h}L_{\ell,h}v_i+u_{\ell,ki,h}M_{\ell,h}v_i\right),\qquad k\ne a,\quad v\in\mathcal{D}^N. \tag*{(17)} ( D a v ) k = v k + N 1 i , ℓ , h ∑ ( u ℓ , ik , h L ℓ , h v i + u ℓ , k i , h M ℓ , h v i ) , k = a , v ∈ D N . (17)
The codomain is V [ N ] ∖ { a } V^{[N]\setminus\{a\}} V [ N ] ∖ { a } . We equip profiles with ∥ v ∥ 2 2 = ∑ i ∥ v i ∥ 2 \lVert v\rVert_2^2=\sum_i\lVert v_i\rVert^2 ∥ v ∥ 2 2 = ∑ i ∥ v i ∥ 2 , and write D ~ a \widetilde{\mathcal{D}}_a D a for the same formula on V N V^N V N . The Hilbert–Schmidt localization from the introduction gives ∑ h ∥ Y ℓ , h ∥ H S 2 = ∥ X ℓ ∥ H S 2 ≤ 2 N \sum_h\lVert Y_{\ell,h}\rVert_{\mathrm{HS}}^2=\lVert X_\ell\rVert_{\mathrm{HS}}^2\le2N ∑ h ∥ Y ℓ , h ∥ HS 2 = ∥ X ℓ ∥ HS 2 ≤ 2 N on Ω N \Omega_N Ω N . Summing the fixed coefficient norms and using ∥ Y ℓ , h ∥ , ∥ Y ~ ℓ , h ∥ ≤ 2 N \lVert Y_{\ell,h}\rVert,\lVert\widetilde{Y}_{\ell,h}\rVert\le\sqrt{2N} ∥ Y ℓ , h ∥ , ∥ Y ℓ , h ∥ ≤ 2 N , we obtain, simultaneously for all a a a ,
∥ D ~ a ∥ ≤ C 0 N . (18) \lVert\widetilde{\mathcal{D}}_a\rVert\le C_0\sqrt{N}. \tag*{(18)} ∥ D a ∥ ≤ C 0 N . (18)
Theorem 3.1 (Block-density dual estimate). If D ≠ 0 \mathcal{D}\ne0 D = 0 , there are E ∗ < ∞ E_*<\infty E ∗ < ∞ and N 0 N_0 N 0 such that, for N ≥ N 0 N\ge N_0 N ≥ N 0 ,
P ( Ω N ∩ { ∃ a ∈ [ N ] : s min ( D a ) ≤ N − E ∗ } ) ≤ C e − c N . (19) \mathbb{P}\left(\Omega_N\cap\left\{\exists a\in[N]:s_{\min}(\mathcal{D}_a)\le N^{-E_*}\right\}\right)\le Ce^{-cN}. \tag*{(19)} P ( Ω N ∩ { ∃ a ∈ [ N ] : s m i n ( D a ) ≤ N − E ∗ } ) ≤ C e − c N . (19) The constants depend only on the fixed spaces, maps, block dimension, and block law. All spaces and dimensions in the proof are real. The second moment ensures q N → 0 q_N\to0 q N → 0 ; the conditional small-ball estimate itself uses bounded densities and a fixed cutoff.
The dimension surplus and the covering geometry
To obtain the dimension surplus from (16) , list the coefficient maps as H 1 , … , H r H_1,\ldots,H_r H 1 , … , H r , where r = 2 n d r=2nd r = 2 n d , and for a nonzero real E ⊂ D E\subset\mathcal{D} E ⊂ D set
Ψ E : E r ⟶ V , Ψ E ( v 1 , … , v r ) = ∑ j H j v j . \Psi_E:E^r\longrightarrow V,\qquad\Psi_E(v_1,\ldots,v_r)=\sum_j H_jv_j. Ψ E : E r ⟶ V , Ψ E ( v 1 , … , v r ) = j ∑ H j v j .
The image of Ψ E \Psi_E Ψ E contains all possible random terms produced by profiles in E N E^N E N . Some of its singular directions may, however, be too small for a useful small-ball estimate. For σ > 0 \sigma>0 σ > 0 , let Y σ ( E ) Y_\sigma(E) Y σ ( E ) be the left singular space corresponding to singular values at least σ \sigma σ . For 0 < τ < 1 0<\tau<1 0 < τ < 1 , let C τ , σ ( E ) C_{\tau,\sigma}(E) C τ , σ ( E ) be the right singular space of P Y σ ( E ) ⊥ ∣ E P_{Y_\sigma(E)^\perp}|_E P Y σ ( E ) ⊥ ∣ E corresponding to singular values at most τ \tau τ . Thus C τ , σ ( E ) C_{\tau,\sigma}(E) C τ , σ ( E ) consists of the directions in E E E that are close to this active image.
These two singular spaces have different roles. Projection onto the active image is where bounded densities give a small-ball estimate. Projection onto its orthogonal complement instead controls the possible profile coordinates: if the random contribution there is small, the profile coordinate must be close to the active image. The second singular space describes the directions that remain free in this restriction. We show that the active image has at least one more dimension, and then construct the corresponding covering net.
The following lemma is the real-space counterpart of [15 Proposition 4.1] . Its surplus of one dimension makes the small-ball estimate dominate the profile net.
Lemma 3.2 (Uniform surplus). There are σ 0 , τ ∈ ( 0 , 1 ) \sigma_0,\tau\in(0,1) σ 0 , τ ∈ ( 0 , 1 ) such that for all nonzero E ⊂ D E\subset\mathcal{D} E ⊂ D and 0 < σ ≤ σ 0 0<\sigma\le\sigma_0 0 < σ ≤ σ 0 ,
y : = dim Y σ ( E ) ≥ c E + 1 , c E : = dim C τ , σ ( E ) . y:=\dim Y_\sigma(E)\ge c_E+1,\qquad c_E:=\dim C_{\tau,\sigma}(E). y := dim Y σ ( E ) ≥ c E + 1 , c E := dim C τ , σ ( E ) . Also ∥ Ψ E ∥ ≥ σ 0 \lVert\Psi_E\rVert\ge\sigma_0 ∥ Ψ E ∥ ≥ σ 0 .
Proof. By (16) and compactness of the finite union of real Grassmannians,
b : = inf 0 ≠ E ⊂ D ∑ j ∥ P E ⊥ H j P E ∥ > 0. b:=\inf_{0\ne E\subset\mathcal{D}}\sum_j\lVert P_{E^\perp}H_jP_E\rVert>0. b := 0 = E ⊂ D inf j ∑ ∥ P E ⊥ H j P E ∥ > 0.
Let M 0 = sup E ∥ Ψ E ∥ < ∞ M_0=\sup_E\lVert\Psi_E\rVert<\infty M 0 = sup E ∥ Ψ E ∥ < ∞ , and write Y = Y σ ( E ) Y=Y_\sigma(E) Y = Y σ ( E ) and C = C τ , σ ( E ) C=C_{\tau,\sigma}(E) C = C τ , σ ( E ) . For x ∈ C x\in C x ∈ C we have
∥ P Y x ∥ 2 ≥ ( 1 − τ 2 ) ∥ x ∥ 2 , \lVert P_Yx\rVert^2\ge(1-\tau^2)\lVert x\rVert^2, ∥ P Y x ∥ 2 ≥ ( 1 − τ 2 ) ∥ x ∥ 2 ,
so P Y : C → Y P_Y:C\to Y P Y : C → Y is injective and c E ≤ y c_E\le y c E ≤ y . Suppose equality holds. Then every w ∈ Y w\in Y w ∈ Y can be written as w = P Y x w=P_Yx w = P Y x for an x ∈ C x\in C x ∈ C , and
dist ( w , E ) ≤ ∥ w − x ∥ ≤ τ ∥ x ∥ ≤ τ 1 − τ 2 ∥ w ∥ . \operatorname{dist}(w,E)\le\lVert w-x\rVert\le\tau\lVert x\rVert\le\frac{\tau}{\sqrt{1-\tau^2}}\lVert w\rVert. dist ( w , E ) ≤ ∥ w − x ∥ ≤ τ ∥ x ∥ ≤ 1 − τ 2 τ ∥ w ∥ .
For a unit v ∈ E v\in E v ∈ E , decompose H j v = P Y H j v + P Y ⊥ H j v H_jv=P_YH_jv+P_{Y^\perp}H_jv H j v = P Y H j v + P Y ⊥ H j v . The second term has norm at most σ \sigma σ , because it belongs to the discarded part of Ψ E \Psi_E Ψ E . Applying the preceding distance bound to the first term gives
dist ( H j v , E ) ≤ M 0 τ 1 − τ 2 + σ . \operatorname{dist}(H_jv,E)\le M_0\frac{\tau}{\sqrt{1-\tau^2}}+\sigma. dist ( H j v , E ) ≤ M 0 1 − τ 2 τ + σ .
Taking operator norms and summing over j j j , we obtain
b ≤ r ( M 0 τ 1 − τ 2 + σ ) . b\le r\left(M_0\frac{\tau}{\sqrt{1-\tau^2}}+\sigma\right). b ≤ r ( M 0 1 − τ 2 τ + σ ) .
For example, take
τ = min ( 1 4 , b 8 r ( 1 + M 0 ) ) , σ 0 = min ( 1 4 , b 4 r ) . \tau=\min\left(\frac14,\frac{b}{8r(1+M_0)}\right),\qquad\sigma_0=\min\left(\frac14,\frac{b}{4r}\right). τ = min ( 4 1 , 8 r ( 1 + M 0 ) b ) , σ 0 = min ( 4 1 , 4 r b ) .
The displayed upper bound is then at most b / 2 b/2 b /2 , a contradiction. Only a positive lower bound for b b b is used. The same argument, or y ≥ 1 y\ge1 y ≥ 1 , gives ∥ Ψ E ∥ ≥ σ 0 \lVert\Psi_E\rVert\ge\sigma_0 ∥ Ψ E ∥ ≥ σ 0 . □
Lemma 3.3 (Real tube net). Let Y = Y σ ( E ) Y=Y_\sigma(E) Y = Y σ ( E ) , and suppose dim E = d \dim E=d dim E = d with c E c_E c E as above. For r t u b e ≥ 0 r_{\mathrm{tube}}\ge0 r tube ≥ 0 , the set { v ∈ E : ∥ v ∥ ≤ 2 , dist ( v , Y ) ≤ r t u b e } \{v\in E:\lVert v\rVert\le2,\ \operatorname{dist}(v,Y)\le r_{\mathrm{tube}}\} { v ∈ E : ∥ v ∥ ≤ 2 , dist ( v , Y ) ≤ r tube } has an h h h -net of size
( C / h ) c E ( C max { r t u b e / τ , h } h ) d − c E , 0 < h ≤ 1. (C/h)^{c_E}\left(\frac{C\max\{r_{\mathrm{tube}}/\tau,h\}}{h}\right)^{d-c_E},\qquad0<h\le1. ( C / h ) c E ( h C max { r tube / τ , h } ) d − c E , 0 < h ≤ 1. The centers may be chosen in the set.
Proof. Write T = P Y ⊥ ∣ E T=P_{Y^\perp}|_E T = P Y ⊥ ∣ E and decompose v = v 0 + v 1 v=v_0+v_1 v = v 0 + v 1 , with v 0 ∈ C τ , σ ( E ) v_0\in C_{\tau,\sigma}(E) v 0 ∈ C τ , σ ( E ) and v 1 ∈ E ∩ C τ , σ ( E ) ⊥ v_1\in E\cap C_{\tau,\sigma}(E)^\perp v 1 ∈ E ∩ C τ , σ ( E ) ⊥ . The singular-vector decomposition gives
∥ T v ∥ 2 = ∥ T v 0 ∥ 2 + ∥ T v 1 ∥ 2 , ∥ T v 1 ∥ ≥ τ ∥ v 1 ∥ . \lVert Tv\rVert^2=\lVert Tv_0\rVert^2+\lVert Tv_1\rVert^2,\qquad\lVert Tv_1\rVert\ge\tau\lVert v_1\rVert. ∥ T v ∥ 2 = ∥ T v 0 ∥ 2 + ∥ T v 1 ∥ 2 , ∥ T v 1 ∥ ≥ τ ∥ v 1 ∥ .
Thus ∥ v 0 ∥ ≤ 2 \lVert v_0\rVert\le2 ∥ v 0 ∥ ≤ 2 and ∥ v 1 ∥ ≤ r t u b e / τ \lVert v_1\rVert\le r_{\mathrm{tube}}/\tau ∥ v 1 ∥ ≤ r tube / τ . Taking Euclidean ball nets in these two orthogonal spaces proves the bound. Finally, moving each net center whose ball meets the set to a point of that intersection changes the mesh by only a fixed factor.
A fixed profile: anchors and independent outputs
The row and column terms share atoms, so independence must be obtained before multiplying output probabilities. We expose all blocks except those joining a small anchor set to the other outputs, and discard the anchor outputs. Each unexposed atom then enters only one remaining equation. The profile values at the anchors must span the active directions. For v ∈ E N v\in E^N v ∈ E N , define its real Gram operator by
Γ v = ∑ i v i ⊗ R v i . \Gamma_v=\sum_i v_i\otimes_{\mathbb{R}}v_i. Γ v = i ∑ v i ⊗ R v i .
Here x ⊗ R x x\otimes_{\mathbb{R}}x x ⊗ R x denotes the map y ↦ ⟨ x , y ⟩ R x y\mapsto\langle x,y\rangle_{\mathbb{R}}x y ↦ ⟨ x , y ⟩ R x . We choose d = dim E d=\dim E d = dim E indices with the following lower bound on their Gram operator.
Lemma 3.4 (Anchor extraction). If 1 ≤ dim E = d ≤ p 1\le\dim E=d\le p 1 ≤ dim E = d ≤ p , γ > 0 \gamma>0 γ > 0 , ∥ v ∥ 2 ≤ C 1 \lVert v\rVert^2\le C_1 ∥ v ∥ 2 ≤ C 1 , and Γ v ⪰ γ P E \Gamma_v\succeq\gamma P_E Γ v ⪰ γ P E , there is an I ⊂ [ N ] I\subset[N] I ⊂ [ N ] of size d d d such that
Γ v , I : = ∑ i ∈ I v i ⊗ R v i ⪰ c d , C 1 γ d N − d P E . \Gamma_{v,I}:=\sum_{i\in I}v_i\otimes_{\mathbb{R}}v_i\succeq c_{d,C_1}\gamma^dN^{-d}P_E. Γ v , I := i ∈ I ∑ v i ⊗ R v i ⪰ c d , C 1 γ d N − d P E .
Proof. Write the profile as a real d d d by N N N matrix B B B in an orthonormal basis of E E E . By Cauchy–Binet,
γ d ≤ det ( B B T ) = ∑ ∣ I ∣ = d det ( B I ) 2 . \gamma^d\le\det(BB^{\mathsf T})=\sum_{|I|=d}\det(B_I)^2. γ d ≤ det ( B B T ) = ∣ I ∣ = d ∑ det ( B I ) 2 .
Choose a maximal minor. Computing determinants on E E E , we obtain det Γ v , I ≥ γ d / ( N d ) \det\Gamma_{v,I}\ge\gamma^d/\binom{N}{d} det Γ v , I ≥ γ d / ( d N ) , whereas every eigenvalue of Γ v , I \Gamma_{v,I} Γ v , I is at most Tr Γ v , I ≤ C 1 \operatorname{Tr}\Gamma_{v,I}\le C_1 Tr Γ v , I ≤ C 1 . Its least eigenvalue is therefore at least C 1 − ( d − 1 ) γ d / ( N d ) C_1^{-(d-1)}\gamma^d/\binom{N}{d} C 1 − ( d − 1 ) γ d / ( d N ) , which proves the lemma.
Bounded joint density alone does not bound the density of a projection from an unbounded set. We therefore restrict the unexposed blocks to a fixed ball before applying the volume estimate below. A small proportion of outputs may fail this restriction; the following event bounds that proportion simultaneously for every possible anchor set. Fix κ = 1 / ( 8 p ) \kappa=1/(8p) κ = 1/ ( 8 p ) and choose a finite R 0 ≥ 1 R_0\ge1 R 0 ≥ 1 so large that
P { ∥ u ∥ ≤ R 0 } 2 n p ≥ 1 − κ / 4. \mathbb{P}\{\lVert u\rVert\le R_0\}^{2np}\ge1-\kappa/4. P {∥ u ∥ ≤ R 0 } 2 n p ≥ 1 − κ /4.
Under the normalization E ∥ u ∥ 2 = 1 \mathbb{E}\lVert u\rVert^2=1 E ∥ u ∥ 2 = 1 , the choice R 0 = 8 p n R_0=8p\sqrt{n} R 0 = 8 p n suffices by Chebyshev’s inequality and a union bound over 2 n p 2np 2 n p whole blocks. More generally, existence of a fixed cutoff requires only that the block law be a probability measure. There is an event A N \mathcal{A}_N A N of probability at least 1 − C e − c N 1-Ce^{-cN} 1 − C e − c N such that, simultaneously for all I ⊂ [ N ] I\subset[N] I ⊂ [ N ] with 1 ≤ ∣ I ∣ ≤ p 1\le|I|\le p 1 ≤ ∣ I ∣ ≤ p , at most κ N \kappa N κ N indices k ∉ I k\notin I k ∈ / I fail
∥ u ℓ , i k ∥ ≤ R 0 , ∥ u ℓ , k i ∥ ≤ R 0 ( i ∈ I , ℓ ∈ [ n ] ) . (20) \|u_{\ell,ik}\|\le R_0,\qquad\|u_{\ell,ki}\|\le R_0\qquad(i\in I,\ \ell\in[n]). \tag*{(20)} ∥ u ℓ , ik ∥ ≤ R 0 , ∥ u ℓ , k i ∥ ≤ R 0 ( i ∈ I , ℓ ∈ [ n ]) . (20)
For a fixed I I I , the indicators in question involve disjoint sets of atoms as k k k varies. Each has failure probability at most κ / 4 \kappa/4 κ /4 , so Chernoff’s inequality bounds the probability of more than κ N \kappa N κ N failures by C e − c N Ce^{-cN} C e − c N . Taking a union over at most p N p pN^p p N p choices of I I I gives the assertion.
Lemma 3.5 (Bounded-support projection). If a random vector Z ∈ R D Z\in\mathbb{R}^{D} Z ∈ R D has density at most K K K and support in a ball of fixed radius R c u t R_{\mathrm{cut}} R cut , and A : R D → R y A:\mathbb{R}^{D}\to\mathbb{R}^{y} A : R D → R y has all y y y singular values at least a > 0 a>0 a > 0 , then, uniformly in w w w ,
P { ∥ A Z − w ∥ ≤ δ } ≤ C D , K , R c u t ( δ / a ) y \mathbb{P}\{\|AZ-w\|\le\delta\}\le C_{D,K,R_{\mathrm{cut}}}(\delta/a)^y P { ∥ A Z − w ∥ ≤ δ } ≤ C D , K , R cut ( δ / a ) y
Proof. In singular-vector coordinates, ∥ A Z − w ∥ ≤ δ \|AZ-w\|\le\delta ∥ A Z − w ∥ ≤ δ restricts the first y y y coordinates to an ellipsoid of volume at most C ( δ / a ) y C(\delta/a)^y C ( δ / a ) y . The other coordinates remain in a fixed ball. The preimage therefore has volume at most C D , R c u t ( δ / a ) y C_{D,R_{\mathrm{cut}}}(\delta/a)^y C D , R cut ( δ / a ) y , and multiplication by the density bound proves the claim. □
We can now combine the anchor lower bound with the cutoff. In the conditioning below each atom remains a whole block, so dependence between its real coordinates does not affect the argument.
Lemma 3.6 (One deterministic profile). In Lemma 3.4 , write y = dim Y σ ( E ) y=\dim Y_\sigma(E) y = dim Y σ ( E ) and b N ( d ) = ⌊ ( 1 − κ ) N ⌋ − d − 1 b_N(d)=\lfloor(1-\kappa)N\rfloor-d-1 b N ( d ) = ⌊( 1 − κ ) N ⌋ − d − 1 . For 0 < δ ≤ 1 0<\delta\le1 0 < δ ≤ 1 , 0 < σ ≤ σ 0 0<\sigma\le\sigma_0 0 < σ ≤ σ 0 , and sufficiently large N N N ,
P { ∥ D a v ∥ 2 ≤ δ , A N } ≤ min { 1 , ( C σ − y γ − d y / 2 N ( d + 1 ) y / 2 δ y ) b N ( d ) } . (21) \mathbb{P}\{\|\mathcal{D}a_v\|_2\le\delta,\ A_N\}\le\min\left\{1,\left(C\sigma^{-y}\gamma^{-dy/2}N^{(d+1)y/2}\delta^y\right)^{b_N(d)}\right\}. \tag*{(21)} P { ∥ D a v ∥ 2 ≤ δ , A N } ≤ min { 1 , ( C σ − y γ − d y /2 N ( d + 1 ) y /2 δ y ) b N ( d ) } . (21) For fixed C 1 C_1 C 1 , the constants and the lower threshold on N N N are uniform over the deterministic profiles, their subspaces, a a a , γ \gamma γ , and σ \sigma σ satisfying these assumptions.
Proof. Choose I I I from Lemma 3.4 for the fixed profile, and let J = [ N ] ∖ ( I ∪ { a } ) J=[N]\setminus(I\cup\{a\}) J = [ N ] ∖ ( I ∪ { a }) . First expose all atoms except the whole directed blocks u ℓ , i k , u ℓ , k i u_{\ell,ik},u_{\ell,ki} u ℓ , ik , u ℓ , k i with i ∈ I i\in I i ∈ I and k ∈ J k\in J k ∈ J . For each output k ∈ J k\in J k ∈ J , its unexposed random part is
Z k = 1 N ∑ i ∈ I , ℓ , h ( u ℓ , i k , h L ℓ , h v i + u ℓ , k i , h M ℓ , h v i ) . Z_k=\frac{1}{\sqrt{N}}\sum_{i\in I,\ell,h}\left(u_{\ell,ik,h}L_{\ell,h}v_i+u_{\ell,ki,h}M_{\ell,h}v_i\right). Z k = N 1 i ∈ I , ℓ , h ∑ ( u ℓ , ik , h L ℓ , h v i + u ℓ , k i , h M ℓ , h v i ) .
The other occurrence of an unexposed atom lies in an anchor output, which is discarded. Consequently the Z k Z_k Z k use disjoint atoms as k k k varies; all other terms are fixed translations. Set Y = Y σ ( E ) Y=Y_\sigma(E) Y = Y σ ( E ) , and let A k A_k A k be the real linear map from the 2 n d d 2ndd 2 n dd coordinates of the unexposed blocks to P Y Z k P_YZ_k P Y Z k . Its Euclidean Gram operator is
1 N P Y ∑ ℓ , h ( L ℓ , h Γ v , I L ℓ , h ∗ + M ℓ , h Γ v , I M ℓ , h ∗ ) P Y ⪰ c σ 2 γ d N − d − 1 I Y . \frac{1}{N}P_Y\sum_{\ell,h}\left(L_{\ell,h}\Gamma_{v,I}L_{\ell,h}^{*}+M_{\ell,h}\Gamma_{v,I}M_{\ell,h}^{*}\right)P_Y\succeq c\sigma^2\gamma^dN^{-d-1}I_Y. N 1 P Y ℓ , h ∑ ( L ℓ , h Γ v , I L ℓ , h ∗ + M ℓ , h Γ v , I M ℓ , h ∗ ) P Y ⪰ c σ 2 γ d N − d − 1 I Y .
Indeed, the anchor lower bound replaces Γ v , I \Gamma_{v,I} Γ v , I by c γ d N − d P E c\gamma^dN^{-d}P_E c γ d N − d P E , and P Y Ψ E Ψ E ∗ P Y ⪰ σ 2 I Y P_Y\Psi_E\Psi_E^*P_Y\succeq\sigma^2I_Y P Y Ψ E Ψ E ∗ P Y ⪰ σ 2 I Y by definition of Y Y Y . Hence every singular value of A k A_k A k onto Y Y Y is at least c σ γ d / 2 N − ( d + 1 ) / 2 c\sigma\gamma^{d/2}N^{-(d+1)/2} c σ γ d /2 N − ( d + 1 ) /2 . This computation uses Euclidean coordinates and real Hilbert adjoints; it requires no assumption on the random covariance inside a block.
Condition on the membership pattern of the good rows (20) in J J J . The rows remain independent, and within each good row the blocks are independent copies of u u u conditioned on ∥ u ∥ ≤ R 0 \|u\|\le R_0 ∥ u ∥ ≤ R 0 . If p 0 = P ( ∥ u ∥ ≤ R 0 ) > 0 p_0=\mathbb{P}(\|u\|\le R_0)>0 p 0 = P ( ∥ u ∥ ≤ R 0 ) > 0 , their product density is at most ( M / p 0 ) 2 n d (M/p_0)^{2nd} ( M / p 0 ) 2 n d on a ball of radius 2 n d R 0 \sqrt{2nd}R_0 2 n d R 0 in R 2 n d \mathbb{R}^{2nd} R 2 n d . The dimension is at most 2 n p d 2npd 2 n p d , so both the density bound and the support radius are uniform in the profile and N N N . Applying Lemma 3.5 to A k A_k A k gives the one-output factor
ϖ = C σ − γ γ − d y / 2 N ( d + 1 ) y / 2 δ y , \varpi= C\sigma^{-\gamma}\gamma^{-d y/2}N^{(d+1)y/2}\delta^y, ϖ = C σ − γ γ − d y /2 N ( d + 1 ) y /2 δ y ,
uniformly in the exposed translations and in the good-row pattern. On A N A_N A N , this particular I I I has at least b N ( d ) b_N(d) b N ( d ) good outputs in J J J . We enlarge to that event, multiply the one-output estimates, and then average over the membership pattern. More explicitly, if F I \mathcal{F}_I F I denotes the first exposure and g g g ranges over patterns with at least b N ( d ) b_N(d) b N ( d ) good rows, the conditional bound is
∑ g P ( g ∣ F I ) min ( 1 , ϖ ) # good ( g ) ≤ min ( 1 , ϖ ) b N ( d ) . \sum_g \mathbb{P}(g \mid\mathcal{F}_I)\min(1,\varpi)^{\#\operatorname{good}(g)}
\leq\min(1,\varpi)^{b_N(d)}. g ∑ P ( g ∣ F I ) min ( 1 , ϖ ) # good ( g ) ≤ min ( 1 , ϖ ) b N ( d ) .
The pattern probabilities sum to at most one. Integrating over F I \mathcal{F}_I F I proves (21) , with A N A_N A N used only through the preceding enlargement, rather than as a conditioning event. □
The estimate for a fixed profile has now been integrated over the exposed atoms. Consequently each point of a deterministic net may use its own anchor set when we sum these probabilities. We construct this net on Ω N \Omega_N Ω N , where the operator norm controls approximation errors; the event Ω N \Omega_N Ω N does not enter any conditional law.
Uniform estimates for regular profiles
We first treat profiles whose Gram operator has a lower bound on their span. Write
R ( E , γ ) = { v ∈ E N : ∥ v ∥ 2 ≤ 1 , Γ v ⪰ γ P E } . \mathcal{R}(E,\gamma)=\{v\in E^N:\|v\|_2\leq1,\ \Gamma_v\succeq\gamma P_E\}. R ( E , γ ) = { v ∈ E N : ∥ v ∥ 2 ≤ 1 , Γ v ⪰ γ P E } .
Norms below one allow the later projection of a unit profile. The Gram lower bound makes anchor extraction uniform. A spectral gap for Ψ E \Psi_E Ψ E bounds its inactive contribution, forcing the profile coordinates into the tubes above. The following lemma assumes both bounds; the next two subsections remove them.
Lemma 3.7 (Real uniform net bound). There are fixed C ∗ , c ∗ , K 0 , L 0 , N ∗ > 0 C_*,c_*,K_0,L_0,N_*>0 C ∗ , c ∗ , K 0 , L 0 , N ∗ > 0 with the following property, uniformly in all the scale parameters below. For definiteness one may take K 0 = L 0 = 2 p 2 + 4 p + 2 K_0=L_0=2p^2+4p+2 K 0 = L 0 = 2 p 2 + 4 p + 2 . Let F ⊂ Gr ( d , D ) F\subset\operatorname{Gr}(d,D) F ⊂ Gr ( d , D ) be the closed family for which all singular values of Ψ E \Psi_E Ψ E lie in [ 0 , η ] ∪ [ σ , ∞ ) [0,\eta]\cup[\sigma,\infty) [ 0 , η ] ∪ [ σ , ∞ ) , where 0 < η < σ ≤ σ 0 0<\eta<\sigma\leq\sigma_0 0 < η < σ ≤ σ 0 . For N ≥ N ∗ N\geq N_* N ≥ N ∗ , if
C ∗ N η ≤ δ , 0 < δ ≤ c ∗ γ ≤ c ∗ , C_*\sqrt{N}\eta\leq\delta,\qquad0<\delta\leq c_*\gamma\leq c_*, C ∗ N η ≤ δ , 0 < δ ≤ c ∗ γ ≤ c ∗ , then
P ( Ω N ∩ A N ∩ { ∃ E ∈ F , v ∈ R ( E , γ ) : ∥ D a v ∥ 2 ≤ δ } ) ≤ C ∗ N N K 0 N γ − p 2 N σ − p N δ N / 2 − L 0 . (22) \mathbb{P}\left(\Omega_N\cap A_N\cap\left\{\exists E\in F,\ v\in\mathcal{R}(E,\gamma):\|\mathcal{D}_av\|_2\leq\delta\right\}\right)
\leq C_*^N N^{K_0N}\gamma^{-p^2N}\sigma^{-pN}\delta^{N/2-L_0}. \tag*{(22)} P ( Ω N ∩ A N ∩ { ∃ E ∈ F , v ∈ R ( E , γ ) : ∥ D a v ∥ 2 ≤ δ } ) ≤ C ∗ N N K 0 N γ − p 2 N σ − pN δ N /2 − L 0 . (22)
Proof. Approximating the subspace. Choose a deterministic net of F F F in projection norm at mesh θ = c δ / N \theta=c\delta/\sqrt{N} θ = cδ / N , with centers in F F F . Orthogonal projections lie in a bounded subset of the real p p p by p p p matrices, so at most ( C / θ ) p 2 (C/\theta)^{p^2} ( C / θ ) p 2 centers are needed. For a nearby center E ′ E' E ′ , define v i 0 = P E ′ v i v_i^0=P_{E'}v_i v i 0 = P E ′ v i . Since the two subspaces have equal dimension and projection distance at most θ \theta θ , we have P E ′ P E P E ′ ⪰ ( 1 − θ 2 ) P E ′ P_{E'}P_EP_{E'}\succeq(1-\theta^2)P_{E'} P E ′ P E P E ′ ⪰ ( 1 − θ 2 ) P E ′ . It follows that
∥ v 0 ∥ 2 ≤ 1 , ∥ v − v 0 ∥ 2 ≤ θ , Γ v 0 ⪰ γ ( 1 − θ 2 ) P E ′ ⪰ ( 15 γ / 16 ) P E ′ . \|v^0\|_2\leq1,\qquad\|v-v^0\|_2\leq\theta,\qquad\Gamma_{v^0}\succeq\gamma(1-\theta^2)P_{E'}\succeq(15\gamma/16)P_{E'}. ∥ v 0 ∥ 2 ≤ 1 , ∥ v − v 0 ∥ 2 ≤ θ , Γ v 0 ⪰ γ ( 1 − θ 2 ) P E ′ ⪰ ( 15 γ /16 ) P E ′ .
On Ω N \Omega_N Ω N , (18) gives ∥ D a v 0 ∥ 2 ≤ 2 δ \|\mathcal{D}_av^0\|_2\leq2\delta ∥ D a v 0 ∥ 2 ≤ 2 δ .
Restricting and covering the profile coordinates. Fix Y = Y σ ( E ′ ) Y = Y_{\sigma}(E') Y = Y σ ( E ′ ) . For each k ≠ a k \ne a k = a , write the random sum in (17) as Ψ E ′ t k \Psi_{E'}t_k Ψ E ′ t k , where the components of t k ∈ ( E ′ ) r t_k \in(E')^r t k ∈ ( E ′ ) r are N − 1 / 2 ∑ i u ℓ , i k , h v i 0 N^{-1/2}\sum_i u_{\ell,ik,h}v_i^0 N − 1/2 ∑ i u ℓ , ik , h v i 0 and N − 1 / 2 ∑ i u ℓ , k i , h v i 0 N^{-1/2}\sum_i u_{\ell,ki,h}v_i^0 N − 1/2 ∑ i u ℓ , k i , h v i 0 . Cauchy–Schwarz gives
∥ t k ∥ 2 ≤ 2 N ∑ ℓ , i , j ∥ u ℓ , i j ∥ 2 ∥ v 0 ∥ 2 2 ≤ 4 n N on Ω N . \lVert t_k\rVert^2 \le\frac{2}{N}\sum_{\ell,i,j}\lVert u_{\ell,ij}\rVert^2\lVert v^0\rVert_2^2 \le4nN \quad\text{on }\Omega_N. ∥ t k ∥ 2 ≤ N 2 ℓ , i , j ∑ ∥ u ℓ , ij ∥ 2 ∥ v 0 ∥ 2 2 ≤ 4 n N on Ω N .
This energy bound uses all coordinates before projection and needs no further exposure of atoms. The spectral gap at the net center implies ∥ P Y ⊥ Ψ E ′ ∥ ≤ η \lVert P_{Y^\perp}\Psi_{E'}\rVert\le\eta ∥ P Y ⊥ Ψ E ′ ∥ ≤ η . Consequently,
dist ( v k 0 , Y ) ≤ ∥ ( D a v 0 ) k ∥ + ∥ P Y ⊥ Ψ E ′ t k ∥ ≤ 2 δ + C N η ≤ C δ . \operatorname{dist}(v_k^0,Y) \le\lVert(\mathcal{D}_a v^0)_k\rVert+\lVert P_{Y^\perp}\Psi_{E'}t_k\rVert\le2\delta+C\sqrt{N}\eta\le C\delta. dist ( v k 0 , Y ) ≤ ∥( D a v 0 ) k ∥ + ∥ P Y ⊥ Ψ E ′ t k ∥ ≤ 2 δ + C N η ≤ C δ .
The small-image equation thus forces each of these profile coordinates into a thin tube about the active image. This restriction is stronger than a ball bound and reduces the number of directions that must be covered at the small mesh. Since the images are taken at the net centers, we do not need continuity of Y σ ( E ) Y_\sigma(E) Y σ ( E ) .
Apply Lemma 3.3 at mesh h = c δ / N h=c\delta/N h = cδ / N to these N − 1 N-1 N − 1 coordinates, and use a ball net for the removed coordinate. If c ′ = dim C τ , σ ( E ′ ) c'=\dim C_{\tau,\sigma}(E') c ′ = dim C τ , σ ( E ′ ) , covering the c ′ c' c ′ tube directions requires ( C N / δ ) c ′ (CN/\delta)^{c'} ( C N / δ ) c ′ per output; the remaining directions have radius O ( δ ) O(\delta) O ( δ ) and require at most ( C N ) p (CN)^p ( C N ) p net points. The product net therefore has size at most
( C N / δ ) c ′ ( N − 1 ) + d ( C N ) p ( N − 1 ) . (23) (CN/\delta)^{c'(N-1)+d}(CN)^{p(N-1)}. \tag*{(23)} ( C N / δ ) c ′ ( N − 1 ) + d ( C N ) p ( N − 1 ) . (23)
Rounding incurs profile error at most c δ / N c\delta/\sqrt{N} cδ / N , so its image error on Ω N \Omega_N Ω N is at most C 0 c δ C_0c\delta C 0 cδ . The inequality
∥ Γ w − Γ v 0 ∥ ≤ ( ∥ w ∥ 2 + ∥ v 0 ∥ 2 ) ∥ w − v 0 ∥ 2 \lVert\Gamma w-\Gamma v^0\rVert\le(\lVert w\rVert_2+\lVert v^0\rVert_2)\lVert w-v^0\rVert_2 ∥ Γ w − Γ v 0 ∥ ≤ (∥ w ∥ 2 + ∥ v 0 ∥ 2 ) ∥ w − v 0 ∥ 2
shows, using δ ≤ c ∗ γ \delta\le c_*\gamma δ ≤ c ∗ γ and taking the fixed mesh constant small enough, that the rounded profile satisfies ∥ w ∥ 2 ≤ 2 \lVert w\rVert_2\le2 ∥ w ∥ 2 ≤ 2 and Γ w ⪰ ( γ / 2 ) P E ′ \Gamma_w\succeq(\gamma/2)P_{E'} Γ w ⪰ ( γ /2 ) P E ′ . Keep only net points with these two properties. Every original profile in the event is represented by one of the selected points, and that point satisfies ∥ D a w ∥ 2 ≤ C δ \lVert\mathcal{D}_a w\rVert_2\le C\delta ∥ D a w ∥ 2 ≤ C δ .
Summing the pointwise estimates. The preceding inclusion holds on Ω N \Omega_N Ω N , so we may drop this event before applying Lemma 3.6 to each deterministic net point. Its factor δ y b N ( d ) \delta^{yb_N(d)} δ y b N ( d ) , combined with the profile net, leaves the exponent
y b N ( d ) − c ′ ( N − 1 ) − d ≥ ( ( 1 − κ ) y − c ′ ) N − y ( d + 2 ) + c ′ − d ≥ ( 7 / 8 ) N − C p , \begin{aligned}
yb_N(d)-c'(N-1)-d\ge\bigl((1-\kappa)y-c'\bigr)N-y(d+2)+c'-d \\
&\ge(7/8)N-C_p,
\end{aligned} y b N ( d ) − c ′ ( N − 1 ) − d ≥ ( ( 1 − κ ) y − c ′ ) N − y ( d + 2 ) + c ′ − d ≥ ( 7/8 ) N − C p ,
because y ≥ c ′ + 1 y\ge c'+1 y ≥ c ′ + 1 , y ≤ p y\le p y ≤ p , and κ = 1 / ( 8 p ) \kappa=1/(8p) κ = 1/ ( 8 p ) . The subspace net contributes the additional fixed power δ − p 2 \delta^{-p^2} δ − p 2 , after which we may weaken the exponent to N / 2 − L 0 N/2-L_0 N /2 − L 0 . The remaining factors are bounded by C ∗ N N K 0 N γ − p 2 N σ − p N C_*^N N^{K_0N}\gamma^{-p^2N}\sigma^{-pN} C ∗ N N K 0 N γ − p 2 N σ − pN , using d , y ≤ p d,y\le p d , y ≤ p and 0 < γ , σ ≤ 1 0<\gamma,\sigma\le1 0 < γ , σ ≤ 1 . More explicitly, their total power of N N N is at most
2 p N + p ( p + 1 ) 2 N + p 2 2 ≤ ( p 2 + 5 p 2 ) N . 2pN+\frac{p(p+1)}{2}N+\frac{p^2}{2}\le\left(p^2+\frac{5p}{2}\right)N. 2 pN + 2 p ( p + 1 ) N + 2 p 2 ≤ ( p 2 + 2 5 p ) N .
The original power of δ \delta δ is at least 7 N / 8 − 2 p 2 − 3 p 7N/8-2p^2-3p 7 N /8 − 2 p 2 − 3 p . These two bounds allow the stated choices K 0 = L 0 = 2 p 2 + 4 p + 2 K_0=L_0=2p^2+4p+2 K 0 = L 0 = 2 p 2 + 4 p + 2 . We take c ∗ c_* c ∗ smaller if necessary so that the radius C δ C\delta C δ used in Lemma 3.6 is at most one. This proves (22) . □ \square □
Removing the spectral gap assumption
The preceding lemma cannot be applied at a single fixed cutoff for Ψ E \Psi_E Ψ E : a singular value could lie just below that cutoff, too small for the density estimate but too large to discard. We instead choose finitely many decreasing scales. Since the target dimension is fixed, one of their intervals is empty, giving the required gap. The recursion below ensures that the inactive contribution at the lower end is smaller than the small-image radius chosen at the upper end.
Lemma 3.8. For every fixed D ≥ 1 D \ge1 D ≥ 1 there is a finite integer e ( D ) e(D) e ( D ) such that, for each 1 ≤ d ≤ dim D 1 \le d \le\dim\mathcal{D} 1 ≤ d ≤ dim D ,
P ( Ω N ∩ A N ∩ { ∃ E ∈ Gr ( d , D ) , v ∈ R ( E , N − D ) : ∥ D a v ∥ 2 ≤ N − e ( D ) } ) ≤ p e − 2 N log N \mathbb{P}\left(\Omega_N \cap\mathcal{A}_N \cap\left\{\exists E \in\operatorname{Gr}(d,\mathcal{D}),\ v \in\mathcal{R}(E,N^{-D}): \|\mathcal{D}a_v\|_2 \le N^{-e(D)}\right\}\right) \le p e^{-2N\log N} P ( Ω N ∩ A N ∩ { ∃ E ∈ Gr ( d , D ) , v ∈ R ( E , N − D ) : ∥ D a v ∥ 2 ≤ N − e ( D ) } ) ≤ p e − 2 N l o g N for all sufficiently large N N N .
Proof. Set
A = 4 p + 8 , B = 4 ( K 0 + p 2 D + 4 ) , f D ( t ) = t A N − B . A = 4p + 8,\qquad B = 4(K_0 + p^2D + 4),\qquad f_D(t) = t^A N^{-B}. A = 4 p + 8 , B = 4 ( K 0 + p 2 D + 4 ) , f D ( t ) = t A N − B .
For N ≥ 4 L 0 N \ge4L_0 N ≥ 4 L 0 , γ = N − D \gamma= N^{-D} γ = N − D and δ = f D ( σ ) \delta= f_D(\sigma) δ = f D ( σ ) , the bound in (22) is at most
( C ∗ N K 0 + p 2 D σ − p δ 1 / 4 ) N = ( C ∗ σ 2 N − 4 ) N ≤ e − 2 N log N . \left(C_*N^{K_0+p^2D}\sigma^{-p}\delta^{1/4}\right)^N = \left(C_*\sigma^2N^{-4}\right)^N \le e^{-2N\log N}. ( C ∗ N K 0 + p 2 D σ − p δ 1/4 ) N = ( C ∗ σ 2 N − 4 ) N ≤ e − 2 N l o g N .
Set t 0 = σ 0 t_0 = \sigma_0 t 0 = σ 0 and t j + 1 = f D ( t j ) N − 3 t_{j+1} = f_D(t_j)N^{-3} t j + 1 = f D ( t j ) N − 3 for 0 ≤ j < p 0 \le j < p 0 ≤ j < p . Lemma 3.2 gives at least one singular value of Ψ E \Psi_E Ψ E at or above t 0 t_0 t 0 . At most p − 1 p-1 p − 1 other positive singular values can therefore lie in the p p p disjoint intervals [ t j + 1 , t j ) [t_{j+1},t_j) [ t j + 1 , t j ) . One interval is empty. All singular values then belong to [ 0 , t j + 1 ] ∪ [ t j , ∞ ) [0,t_{j+1}] \cup[t_j,\infty) [ 0 , t j + 1 ] ∪ [ t j , ∞ ) , so this subspace is in the gap family of Lemma 3.7 with σ = t j \sigma= t_j σ = t j , η = t j + 1 \eta= t_{j+1} η = t j + 1 , and δ j = f D ( t j ) \delta_j = f_D(t_j) δ j = f D ( t j ) . The contribution of the inactive singular directions is bounded by
C ∗ N t j + 1 = C ∗ N − 5 / 2 δ j ≤ δ j . C_*\sqrt{N}\,t_{j+1} = C_*N^{-5/2}\delta_j \le\delta_j. C ∗ N t j + 1 = C ∗ N − 5/2 δ j ≤ δ j .
Also δ j ≤ c ∗ N − D \delta_j \le c_*N^{-D} δ j ≤ c ∗ N − D for large N N N , since B > D B > D B > D .
The recursion gives
t j = σ 0 A j N − ( B + 3 ) ( A j − 1 ) / ( A − 1 ) . t_j = \sigma_0^{A^j}N^{-(B+3)(A^j-1)/(A-1)}. t j = σ 0 A j N − ( B + 3 ) ( A j − 1 ) / ( A − 1 ) .
Any integer e ( D ) e(D) e ( D ) greater than A ( B + 3 ) ( A p − 1 − 1 ) / ( A − 1 ) + B + 2 A(B+3)(A^{p-1}-1)/(A-1)+B+2 A ( B + 3 ) ( A p − 1 − 1 ) / ( A − 1 ) + B + 2 ensures N − e ( D ) ≤ min j < p δ j N^{-e(D)} \le\min_{j<p}\delta_j N − e ( D ) ≤ min j < p δ j for all sufficiently large N N N . A union over the p p p gap families proves the lemma. □
Projection to profiles with a Gram lower bound
The spectral gap has been removed, but the last lemma still assumes a polynomial lower bound for the profile Gram operator on its span. An arbitrary unit profile need not satisfy such a bound. We diagonalize its Gram operator and discard the directions with small eigenvalues, projecting every component onto the same remaining subspace. The squared error is the sum of the discarded eigenvalues. The thresholds below make this error small enough, even after multiplication by the localized operator norm, to apply the preceding estimate to the projected profile.
Proof of Theorem 3.1. Let q = dim D q = \dim\mathcal{D} q = dim D . Choose integers
D 1 = 1 , D d + 1 ≥ max { D d + 1 , 2 e ( D d ) + 4 } , 1 ≤ d < q . D_1 = 1,\qquad D_{d+1} \ge\max\{D_d + 1, 2e(D_d) + 4\},\qquad1 \le d < q. D 1 = 1 , D d + 1 ≥ max { D d + 1 , 2 e ( D d ) + 4 } , 1 ≤ d < q .
For a unit profile v ∈ D N v \in\mathcal{D}^N v ∈ D N , diagonalize its real Gram operator. Starting at dimension q q q , discard its smallest eigenvector whenever the smallest remaining eigenvalue is below N − D d N^{-D_d} N − D d at dimension d d d . Project every profile component at each step, without renormalizing. The process stops at some d ≥ 1 d \ge1 d ≥ 1 , because the largest original eigenvalue is at least 1 / q > N − 1 1/q > N^{-1} 1/ q > N − 1 . Let E E E be the remaining eigenspace and v ′ v' v ′ the projected profile. Then v ′ ∈ R ( E , N − D d ) v' \in\mathcal{R}(E,N^{-D_d}) v ′ ∈ R ( E , N − D d ) . At a discarded dimension j > d j>d j > d , the squared projection error is the discarded eigenvalue, which is less than N − D j N^{-D_j} N − D j . These errors are orthogonal, so
∥ v − v ′ ∥ 2 2 ≤ ∑ j = d + 1 q N − D j ≤ q N − D d + 1 ( d < q ) . \lVert v-v'\rVert_2^2 \le\sum_{j=d+1}^{q}N^{-D_j}\le qN^{-D_{d+1}}\qquad(d<q). ∥ v − v ′ ∥ 2 2 ≤ j = d + 1 ∑ q N − D j ≤ q N − D d + 1 ( d < q ) .
There is no error when d = q d=q d = q . On Ω N \Omega_N Ω N , (18) and the choice of D d + 1 D_{d+1} D d + 1 now give
∥ D a ( v − v ′ ) ∥ 2 ≤ C 0 q N N − D d + 1 / 2 ≤ 1 2 N − e ( D d ) \lVert\mathcal{D}_a(v-v')\rVert_2\le C_0\sqrt{qN}\,N^{-D_{d+1}/2}\le\frac{1}{2}N^{-e(D_d)} ∥ D a ( v − v ′ ) ∥ 2 ≤ C 0 q N N − D d + 1 /2 ≤ 2 1 N − e ( D d )
for all sufficiently large N N N . Choose E ∗ > max d e ( D d ) + 1 E_*>\max_d e(D_d)+1 E ∗ > max d e ( D d ) + 1 . If ∥ D a v ∥ 2 ≤ N − E ∗ \lVert\mathcal{D}_av\rVert_2\le N^{-E_*} ∥ D a v ∥ 2 ≤ N − E ∗ , the triangle inequality then gives ∥ D a v ′ ∥ 2 ≤ N − e ( D d ) \lVert\mathcal{D}_av'\rVert_2\le N^{-e(D_d)} ∥ D a v ′ ∥ 2 ≤ N − e ( D d ) . Lemma 3.8 bounds this event, within Ω N ∩ A N \Omega_N\cap A_N Ω N ∩ A N , by p e − 2 N log N pe^{-2N\log N} p e − 2 N l o g N for each possible d d d and a a a . Taking the union over them and adding P ( A N c ) ≤ C e − c N \mathbb{P}(A_N^c)\le Ce^{-cN} P ( A N c ) ≤ C e − c N once proves (19) .
The families with a spectral gap and the profile sets with a Gram lower bound used above are compact and defined by continuous maps. Their existential events are therefore measurable, as is the least singular value, which is continuous in the entries. □
From inverse rows to the least singular value
Assume that the coefficient tuple is nonzero and complex irreducible. The trace observations of § 2 control an inverse row because its equations prevent concentration in the kernel of O \mathcal{O} O . We sum these row bounds to control the full inverse, then identify the paired traces with logarithmic derivatives of the full determinant. Relative anti-concentration gives the least singular value bound.
The inverse equations and a frame estimate
On the invertibility set, write R ‾ = T ‾ N − 1 \overline{R}=\overline{T}_N^{-1} R = T N − 1 . For a ∈ [ N ] a\in[N] a ∈ [ N ] , set Φ i = R a i ∗ ∈ V \Phi_i=R_{ai}^*\in V Φ i = R ai ∗ ∈ V . The ( a , k ) (a,k) ( a , k ) block of R ‾ T ‾ N = I \overline{R}\overline{T}_N=I R T N = I vanishes when k ≠ a k\ne a k = a . Taking its adjoint gives
Φ k + 1 N ∑ i , ℓ , h ( u ℓ , i k , h C ℓ , h , 1 ∗ Φ i + u ℓ , k i , h C ℓ , h , 2 ∗ Φ i ) = 0 , k ≠ a . (24) \Phi_k+\frac{1}{\sqrt{N}}\sum_{i,\ell,h}\left(u_{\ell,ik,h}C_{\ell,h,1}^*\Phi_i+u_{\ell,ki,h}C_{\ell,h,2}^*\Phi_i\right)=0,\qquad k\ne a. \tag*{(24)} Φ k + N 1 i , ℓ , h ∑ ( u ℓ , ik , h C ℓ , h , 1 ∗ Φ i + u ℓ , k i , h C ℓ , h , 2 ∗ Φ i ) = 0 , k = a . (24)
Left multiplication by C ℓ , h , 1 ∗ C_{\ell,h,1}^* C ℓ , h , 1 ∗ and C ℓ , h , 2 ∗ C_{\ell,h,2}^* C ℓ , h , 2 ∗ preserves V V V . Set L ℓ , h Φ = C ℓ , h , 1 ∗ Φ L_{\ell,h}\Phi=C_{\ell,h,1}^*\Phi L ℓ , h Φ = C ℓ , h , 1 ∗ Φ and M_{\ell,h}\Phi=C_{\ell,h,2}^*\Phi. By Lemma 2.5, Theorem 3.1 applies with \mathcal{D}=\ker\mathcal{O}$ to the component of the profile in D \mathcal{D} D .
Proposition 4.1 (Restricted normal-space frame). There are β < ∞ \beta<\infty β < ∞ and an event E N ⊂ Ω N \mathcal{E}_N\subset\Omega_N E N ⊂ Ω N such that P ( Ω N ∖ E N ) ≤ C e − c N \mathbb{P}(\Omega_N\setminus\mathcal{E}_N)\le Ce^{-cN} P ( Ω N ∖ E N ) ≤ C e − c N and, on E N \mathcal{E}_N E N , for every a a a and every profile Φ ∈ V N \Phi\in V^N Φ ∈ V N satisfying (24) ,
1 N ∑ i , ℓ , h , t ∣ Tr ( Φ i ∗ C ℓ , h , t ) ∣ 2 ≥ N − β ∑ i ∥ Φ i ∥ H S 2 . (25) \frac{1}{N}\sum_{i,\ell,h,t}\left\lvert\operatorname{Tr}(\Phi_i^*C_{\ell,h,t})\right\rvert^2\ge N^{-\beta}\sum_i\lVert\Phi_i\rVert_{\mathrm{HS}}^2. \tag*{(25)} N 1 i , ℓ , h , t ∑ ∣ Tr ( Φ i ∗ C ℓ , h , t ) ∣ 2 ≥ N − β i ∑ ∥ Φ i ∥ HS 2 . (25)
Proof. Use real orthogonal projections in V V V to write Φ i = v i + w i \Phi_i=v_i+w_i Φ i = v i + w i , where v i ∈ D v_i\in\mathcal{D} v i ∈ D and w i ∈ D ⊥ w_i\in\mathcal{D}^{\perp} w i ∈ D ⊥ . Equation (24) says
D a v = − D ~ a w . \mathcal{D}_av=-\widetilde{\mathcal{D}}_aw. D a v = − D a w .
Here D a \mathcal{D}_a D a is the restriction to D N \mathcal{D}^N D N , while D ~ a \widetilde{\mathcal{D}}_a D a denotes the same expression on V N V^N V N . Theorem 3.1 and the localized operator norm bound give, simultaneously for all a a a ,
N − E ∗ ∥ v ∥ 2 ≤ ∥ D a v ∥ 2 = ∥ D ~ a w ∥ 2 ≤ C 0 N ∥ w ∥ 2 . N^{-E_*}\lVert v\rVert_2\le\lVert\mathcal{D}_av\rVert_2=\lVert\widetilde{\mathcal{D}}_aw\rVert_2\le C_0\sqrt{N}\,\lVert w\rVert_2. N − E ∗ ∥ v ∥ 2 ≤ ∥ D a v ∥ 2 = ∥ D a w ∥ 2 ≤ C 0 N ∥ w ∥ 2 .
Consequently,
∥ Φ ∥ 2 2 = ∥ v ∥ 2 2 + ∥ w ∥ 2 2 ≤ ( 1 + C 0 2 N 2 E ∗ + 1 ) ∥ w ∥ 2 2 . \lVert\Phi\rVert_2^2=\lVert v\rVert_2^2+\lVert w\rVert_2^2\le\left(1+C_0^2N^{2E_*+1}\right)\lVert w\rVert_2^2. ∥ Φ ∥ 2 2 = ∥ v ∥ 2 2 + ∥ w ∥ 2 2 ≤ ( 1 + C 0 2 N 2 E ∗ + 1 ) ∥ w ∥ 2 2 .
The restriction of O \mathcal{O} O to D ⊥ D^\perp D ⊥ is injective. By finite dimensionality there is a constant λ > 0 \lambda>0 λ > 0 such that ∥ O ( w ) ∥ 2 2 ≥ λ ∥ w ∥ 2 \|\mathcal{O}(w)\|_2^2\geq\lambda\|w\|^2 ∥ O ( w ) ∥ 2 2 ≥ λ ∥ w ∥ 2 there. Using O ( Φ i ) = O ( w i ) \mathcal{O}(\Phi_i)=\mathcal{O}(w_i) O ( Φ i ) = O ( w i ) , we obtain
1 N ∑ i ∥ O ( Φ i ) ∥ 2 2 ≥ λ N ∥ w ∥ 2 2 ≥ c N − 2 E ∗ − 2 ∥ Φ ∥ 2 2 . \frac{1}{N}\sum_i\|\mathcal{O}(\Phi_i)\|_2^2\geq\frac{\lambda}{N}\|w\|_2^2\geq cN^{-2E_*-2}\|\Phi\|_2^2. N 1 i ∑ ∥ O ( Φ i ) ∥ 2 2 ≥ N λ ∥ w ∥ 2 2 ≥ c N − 2 E ∗ − 2 ∥Φ ∥ 2 2 .
Take β = 2 E ∗ + 3 \beta=2E_*+3 β = 2 E ∗ + 3 and enlarge N 0 N_0 N 0 to absorb c c c . The exceptional probability is the one in Theorem 3.1. If D = 0 D=0 D = 0 , injectivity alone proves the result with E N = Ω N \mathcal{E}_N=\Omega_N E N = Ω N and, for example, β = 2 \beta=2 β = 2 .
The full determinant gradient
Jacobi’s formula identifies the normalized derivatives of det T N \det T_N det T N with the paired traces. Summing Proposition 4.1 over the inverse rows therefore gives a bound through the full determinant gradient. In [15 Sections 3.4–3.7 and 6.1] , the determinant is instead that of a projected exposed block of fixed size. The present summation accommodates the atoms shared by different columns.
Lemma 4.2 (The determinant gradient in atom coordinates). Let F = det T N F=\det T_N F = det T N as a polynomial in the n d N 2 ndN^2 n d N 2 original real coordinates u ℓ , i j , h u_{\ell,ij,h} u ℓ , ij , h . It has F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 , degree at most 2 m 2m 2 m in each coordinate, and, with R = T N − 1 R=T_N^{-1} R = T N − 1 ,
∂ u ℓ , i j , h F F = 1 N ( Tr ( R j i A ℓ , h ) + Tr ( R i j B ℓ , h ) ) . (26) \frac{\partial_{u_{\ell,ij,h}}F}{F}=\frac{1}{\sqrt{N}}\left(\operatorname{Tr}(R_{ji}A_{\ell,h})+\operatorname{Tr}(R_{ij}B_{\ell,h})\right). \tag*{(26)} F ∂ u ℓ , ij , h F = N 1 ( Tr ( R j i A ℓ , h ) + Tr ( R ij B ℓ , h ) ) . (26) On E N \mathcal{E}_N E N from Proposition 4.1 , after removing a null set,
∥ R ∥ ≤ ∥ R ∥ H S ≤ N β / 2 ∥ ∇ F ∥ 2 ∣ F ∣ . (27) \|R\|\leq\|R\|_{\mathrm{HS}}\leq N^{\beta/2}\frac{\|\nabla F\|_2}{|F|}. \tag*{(27)} ∥ R ∥ ≤ ∥ R ∥ HS ≤ N β /2 ∣ F ∣ ∥∇ F ∥ 2 . (27)
Proof. The derivative matrix in the coordinate u ℓ , i j , h u_{\ell,ij,h} u ℓ , ij , h is
∂ u ℓ , i j , h T N = N − 1 / 2 ( A ℓ , h ⊗ E i j + B ℓ , h ⊗ E j i ) . \partial_{u_{\ell,ij,h}}T_N=N^{-1/2}(A_{\ell,h}\otimes E_{ij}+B_{\ell,h}\otimes E_{ji}). ∂ u ℓ , ij , h T N = N − 1/2 ( A ℓ , h ⊗ E ij + B ℓ , h ⊗ E j i ) .
Its rank is at most 2 m 2m 2 m , so determinant multilinearity gives the coordinate degree bound. Jacobi’s formula then yields
∂ u ℓ , i j , h F F = Tr ( R ∂ u ℓ , i j , h T N ) = N − 1 / 2 ( Tr ( R j i A ℓ , h ) + Tr ( R i j B ℓ , h ) ) . \frac{\partial_{u_{\ell,ij,h}}F}{F}=\operatorname{Tr}(R\,\partial_{u_{\ell,ij,h}}T_N)=N^{-1/2}\left(\operatorname{Tr}(R_{ji}A_{\ell,h})+\operatorname{Tr}(R_{ij}B_{\ell,h})\right). F ∂ u ℓ , ij , h F = Tr ( R ∂ u ℓ , ij , h T N ) = N − 1/2 ( Tr ( R j i A ℓ , h ) + Tr ( R ij B ℓ , h ) ) .
When i = j i=j i = j , the derivative matrix contains A ℓ , h + B ℓ , h A_{\ell,h}+B_{\ell,h} A ℓ , h + B ℓ , h in that one block, so the same formula applies without an additional diagonal factor. In class (B) the two derivatives are
∂ ℜ ξ ℓ , i j F F = N − 1 / 2 ( Tr ( R j i A ℓ ) + Tr ( R i j B ℓ ) ) , \frac{\partial_{\Re\xi_{\ell,ij}}F}{F}=N^{-1/2}\left(\operatorname{Tr}(R_{ji}A_\ell)+\operatorname{Tr}(R_{ij}B_\ell)\right), F ∂ ℜ ξ ℓ , ij F = N − 1/2 ( Tr ( R j i A ℓ ) + Tr ( R ij B ℓ ) ) ,
∂ ℑ ξ ℓ , i j F F = i N − 1 / 2 ( Tr ( R j i A ℓ ) − Tr ( R i j B ℓ ) ) . \frac{\partial_{\Im\xi_{\ell,ij}}F}{F}=iN^{-1/2}\left(\operatorname{Tr}(R_{ji}A_\ell)-\operatorname{Tr}(R_{ij}B_\ell)\right). F ∂ ℑ ξ ℓ , ij F = i N − 1/2 ( Tr ( R j i A ℓ ) − Tr ( R ij B ℓ ) ) .
These identities hold whether or not the two atom coordinates are independent. Since F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 , its zero set is Lebesgue null in that real coordinate space. The product block law is absolutely continuous, so T N T_N T N is invertible almost surely and { F = 0 } \{F=0\} { F = 0 } can be removed from E N \mathcal{E}_N E N without changing its probability.
To control the full inverse, apply (25) to Φ i = R ^ a i ∗ \Phi_i=\widehat{R}_{ai}^* Φ i = R ai ∗ and sum over a a a . Equations (13) –(14) give
∑ a , i , ℓ , h , t ∣ Tr ( R ^ a i C ℓ , h , t ) ∣ 2 = 2 N ∥ ∇ F ∥ 2 2 ∣ F ∣ 2 , ∑ a , i ∥ R ^ a i ∥ H S 2 = 2 ∥ R ∥ H S 2 . \sum_{a,i,\ell,h,t}\left|\operatorname{Tr}(\widehat{R}_{ai}C_{\ell,h,t})\right|^2=2N\frac{\|\nabla F\|_2^2}{|F|^2},\qquad\sum_{a,i}\|\widehat{R}_{ai}\|_{\mathrm{HS}}^2=2\|R\|_{\mathrm{HS}}^2. a , i , ℓ , h , t ∑ Tr ( R ai C ℓ , h , t ) 2 = 2 N ∣ F ∣ 2 ∥∇ F ∥ 2 2 , a , i ∑ ∥ R ai ∥ HS 2 = 2∥ R ∥ HS 2 .
For each ( ℓ , h ) (\ell,h) ( ℓ , h ) , the t = 1 t=1 t = 1 term at ( a , i ) (a,i) ( a , i ) is N \sqrt{N} N times ∂ u ℓ , i a , h F / F \partial_{u_{\ell,ia,h}}F/F ∂ u ℓ , ia , h F / F , and the t = 2 t=2 t = 2 term is N \sqrt{N} N times ∂ u ℓ , a i , h F / F \partial_{u_{\ell,ai,h}}F/F ∂ u ℓ , ai , h F / F . Thus each derivative, including a diagonal derivative, occurs twice in the full sum. The squared inverse norm is also doubled by (12) . Substituting the two displayed identities into the sum of (25) gives
2 ∥ ∇ F ∥ 2 2 ∣ F ∣ 2 ≥ 2 N − β ∥ R ∥ H S 2 2\frac{\lVert\nabla F\rVert_2^2}{\lvert F\rvert^2}\geq2N^{-\beta}\lVert R\rVert_{\mathrm{HS}}^2 2 ∣ F ∣ 2 ∥ ∇ F ∥ 2 2 ≥ 2 N − β ∥ R ∥ HS 2
which proves (27) for F = det T N F=\det T_N F = det T N .
Relative polynomial anti-concentration
The inverse bound reduces the least singular value event to a determinant that is small relative to its gradient. We bound this relative event directly under the atom law.
Lemma 4.3 (Relative polynomial anti-concentration). Let F F F be a nonzero complex-valued polynomial in real coordinates, of degree at most d d d in each coordinate.
(i) Let X 1 , … , X k X_1,\ldots,X_k X 1 , … , X k be independent real random variables with densities f j f_j f j satisfying ∥ f j ∥ L ∞ ( R ) ≤ M \lVert f_j\rVert_{L^\infty(\mathbb{R})}\leq M ∥ f j ∥ L ∞ ( R ) ≤ M . Writing X = ( X 1 , … , X k ) X=(X_1,\ldots,X_k) X = ( X 1 , … , X k ) , we have
P { ∣ F ( X ) ∣ ≤ t ∥ ∇ F ( X ) ∥ 2 } ≤ 2 M d 2 k 3 / 2 t , t > 0. (28) \mathbb{P}\{\lvert F(X)\rvert\leq t\lVert\nabla F(X)\rVert_2\}\leq2Md^2k^{3/2}t,\qquad t>0. \tag*{(28)} P {∣ F ( X )∣ ≤ t ∥ ∇ F ( X ) ∥ 2 } ≤ 2 M d 2 k 3/2 t , t > 0. (28) (ii) For k k k independent R 2 \mathbb{R}^2 R 2 -valued blocks, each with joint density bounded by M M M and second moment at most M 2 M_2 M 2 ,
P { ∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C d , M , M 2 k 3 / 2 t 1 / 2 , 0 < t ≤ 1. (29) \mathbb{P}\{\lvert F\rvert\leq t\lVert\nabla F\rVert_2\}\leq C_{d,M,M_2}k^{3/2}t^{1/2},\qquad0<t\leq1. \tag*{(29)} P {∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C d , M , M 2 k 3/2 t 1/2 , 0 < t ≤ 1. (29) Here the gradient contains all 2 k 2k 2 k real coordinate derivatives.
Proof. For a nonzero univariate polynomial q q q of degree b ≤ d b\leq d b ≤ d , q ′ / q = ∑ j = 1 b ( x − z j ) − 1 q'/q=\sum_{j=1}^{b}(x-z_j)^{-1} q ′ / q = ∑ j = 1 b ( x − z j ) − 1 off its zeros. Thus ∣ q ∣ ≤ v ∣ q ′ ∣ \lvert q\rvert\leq v\lvert q'\rvert ∣ q ∣ ≤ v ∣ q ′ ∣ forces ∣ x − z j ∣ ≤ b v \lvert x-z_j\rvert\leq bv ∣ x − z j ∣ ≤ b v for some root. On the real line these intervals have total length at most 2 d 2 v 2d^2v 2 d 2 v . A nonzero constant gives an empty event, and a finite zero set does not affect length.
For (i), on the event with F ≠ 0 F\neq0 F = 0 some coordinate satisfies ∣ F ∣ ≤ t k ∣ ∂ j F ∣ \lvert F\rvert\leq t\sqrt{k}\lvert\partial_jF\rvert ∣ F ∣ ≤ t k ∣ ∂ j F ∣ . Condition on all other coordinates and apply the length estimate with v = t k v=t\sqrt{k} v = t k . Independence leaves the selected coordinate density at most M M M . Sum over k k k coordinates to obtain (28) .
For (ii), fix a block ( x , y ) (x,y) ( x , y ) and condition on all the other blocks. The resulting polynomial in ( x , y ) (x,y) ( x , y ) is nonzero for almost every choice of the other blocks: in the expansion in powers of x , y x,y x , y , at least one coefficient is a nonzero polynomial of those blocks. For a nonzero bivariate polynomial, the exceptional y y y for which the x x x -slice vanishes identically form a finite set: at least one coefficient polynomial in y y y is nonzero. All other slices obey the univariate length estimate; a nonzero constant slice contributes nothing. For a cutoff R c u t > 0 R_{\mathrm{cut}}>0 R cut > 0 , Fubini gives
Leb 2 { x 2 + y 2 ≤ R c u t 2 : ∣ F ( x , y ) ∣ ≤ v ∣ ∂ x F ( x , y ) ∣ } ≤ 4 d 2 R c u t v . \operatorname{Leb}_2\{x^2+y^2\leq R_{\mathrm{cut}}^2:\lvert F(x,y)\rvert\leq v\lvert\partial_xF(x,y)\rvert\}\leq4d^2R_{\mathrm{cut}}v. Leb 2 { x 2 + y 2 ≤ R cut 2 : ∣ F ( x , y )∣ ≤ v ∣ ∂ x F ( x , y )∣} ≤ 4 d 2 R cut v .
The same estimate holds for ∂ y F \partial_yF ∂ y F . Multiplying the area by the joint density bound M M M bounds the conditional probability. The coordinates x x x and y y y may be dependent, since we conditioned only on the other blocks.
If all k k k blocks have norm at most R c u t R_{\mathrm{cut}} R cut , the event in (ii) implies one of the 2 k 2k 2 k coordinate events with v = t 2 k v=t\sqrt{2k} v = t 2 k . The complement has probability at most k M 2 / R c u t 2 kM_2/R_{\mathrm{cut}}^2 k M 2 / R cut 2 . Therefore
P { ∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C M d 2 k 3 / 2 R c u t t + k M 2 R c u t − 2 . \mathbb{P}\{\lvert F\rvert\leq t\lVert\nabla F\rVert_2\}\leq CMd^2k^{3/2}R_{\mathrm{cut}}t+kM_2R_{\mathrm{cut}}^{-2}. P {∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C M d 2 k 3/2 R cut t + k M 2 R cut − 2 .
Take R c u t = t − 1 / 2 R_{\mathrm{cut}}=t^{-1/2} R cut = t − 1/2 for 0 < t ≤ 1 0<t\le1 0 < t ≤ 1 ; then R c u t t = t 1 / 2 R_{\mathrm{cut}}t=t^{1/2} R cut t = t 1/2 and t ≤ t 1 / 2 t\le t^{1/2} t ≤ t 1/2 , proving (29) . In both parts, all exceptional coefficient zero sets and { F = 0 } \{F=0\} { F = 0 } are null under the product of the absolutely continuous block laws.
The least singular value cutoff
Proposition 4.4 (Irreducible pencils in both input classes). For every nonzero complex irreducible tuple,
P { s min ( T N ) ≤ t , Ω N } ≤ C N K t ω + C e − c N . \mathbb{P}\{s_{\min}(T_N)\le t,\ \Omega_N\}\le C N^K t^\omega+C e^{-cN}. P { s m i n ( T N ) ≤ t , Ω N } ≤ C N K t ω + C e − c N .
For the Gaussian laws in Theorem 1.2 , one may take ω = 1 \omega=1 ω = 1 also in class (B).
Proof. On E N \mathcal{E}_N E N , the event of a small least singular value implies ∣ F ∣ ≤ t N β / 2 ∥ ∇ F ∥ 2 |F|\le tN^{\beta/2}\|\nabla F\|_2 ∣ F ∣ ≤ t N β /2 ∥∇ F ∥ 2 . In class (A), apply Lemma 4.3 (i) with d = 2 m d=2m d = 2 m and k = n N 2 k=nN^2 k = n N 2 to get C N 3 + β / 2 t CN^{3+\beta/2}t C N 3 + β /2 t . In class (B), set v = t N β / 2 v=tN^{\beta/2} v = t N β /2 and use Lemma 4.3 (ii) with k = n N 2 k=nN^2 k = n N 2 , d = 2 m d=2m d = 2 m , M 2 = 1 M_2=1 M 2 = 1 . For v ≤ 1 v\le1 v ≤ 1 this gives C N 3 + β / 4 t 1 / 2 CN^{3+\beta/4}t^{1/2} C N 3 + β /4 t 1/2 . For v > 1 v>1 v > 1 the same upper bound, enlarged if necessary, is at least one. Adding P ( Ω N ∖ E N ) \mathbb{P}(\Omega_N\setminus\mathcal{E}_N) P ( Ω N ∖ E N ) proves the bound.
For the Gaussian refinement in class (B), enumerate the n N 2 nN^2 n N 2 atom blocks by j j j and write each block as u = S g u=Sg u = S g , where g g g has two independent standard real Gaussian coordinates and the fixed 2 × 2 2\times2 2 × 2 real matrix S S S is invertible. Let F ~ ( g ) = F ( ( S g j ) j ) \widetilde{F}(g)=F((Sg_j)_j) F ( g ) = F (( S g j ) j ) . By the chain rule,
∥ ∇ g F ~ ∥ 2 ≥ s min ( S ) ∥ ∇ u F ∥ 2 . \|\nabla_g\widetilde{F}\|_2\ge s_{\min}(S)\|\nabla_uF\|_2. ∥ ∇ g F ∥ 2 ≥ s m i n ( S ) ∥ ∇ u F ∥ 2 .
Varying one coordinate of g j g_j g j varies just one complex entry by a fixed complex multiple and its conjugate entry in X ∗ X^* X ∗ by the conjugate multiple. The matrix variation still has rank at most 2 m 2m 2 m . Thus F ~ \widetilde{F} F has degree at most 2 m 2m 2 m in each of its 2 n N 2 2nN^2 2 n N 2 independent real Gaussian coordinates. Apply Lemma 4.3 (i) to F ~ \widetilde{F} F with radius t N β / 2 / s min ( S ) tN^{\beta/2}/s_{\min}(S) t N β /2 / s m i n ( S ) . The resulting bound C N 3 + β / 2 t CN^{3+\beta/2}t C N 3 + β /2 t therefore bounds the least singular value event on E N \mathcal{E}_N E N .
The next section combines the factor bounds. The resulting cutoff enters the lower logarithmic tail estimate, together with a count of intermediate small singular values.
Reconstruction from irreducible factors
The inverse estimate applies to every nonzero irreducible factor. The passage to arbitrary coefficients follows [15 Section 3.1] : a composition series makes the pencil upper triangular, and a finite inverse expansion combines its diagonal bounds. Localization controls the off-diagonal blocks.
Proof of Theorem 1.2 . Choose a composition series of the natural C m \mathbb{C}^m C m -module for the original complex coefficient algebra, and an orthogonal splitting adapted to that series. Every coefficient matrix is then block upper triangular, with r c o m p ≤ m r_{\mathrm{comp}}\le m r comp ≤ m diagonal factors. The nonzero factors are complex irreducible; a factor on which every coefficient acts as zero contributes an identity pencil.
Write T = T N = D + J T=T_N=D+J T = T N = D + J , with D = diag ( D 1 , … , D r c o m p ) D=\operatorname{diag}(D_1,\ldots,D_{r_{\mathrm{comp}}}) D = diag ( D 1 , … , D r comp ) block diagonal and J J J strictly block upper triangular in these r c o m p r_{\mathrm{comp}} r comp factors. On Ω N \Omega_N Ω N , ∥ J ∥ ≤ C N \|J\|\le C\sqrt{N} ∥ J ∥ ≤ C N . Suppose that each diagonal factor has least singular value at least τ f a c ≤ 1 \tau_{\mathrm{fac}}\le1 τ fac ≤ 1 , so that ∥ D − 1 ∥ ≤ τ f a c − 1 \|D^{-1}\|\le\tau_{\mathrm{fac}}^{-1} ∥ D − 1 ∥ ≤ τ fac − 1 . The matrix D − 1 J D^{-1}J D − 1 J is strictly block upper triangular and ( D − 1 J ) r c o m p = 0 (D^{-1}J)^{r_{\mathrm{comp}}}=0 ( D − 1 J ) r comp = 0 . Therefore
T − 1 = ( I + D − 1 J ) − 1 D − 1 = ∑ j = 0 r c o m p − 1 ( − D − 1 J ) j D − 1 . (30) T^{-1}=(I+D^{-1}J)^{-1}D^{-1}
=\sum_{j=0}^{r_{\mathrm{comp}}-1}(-D^{-1}J)^jD^{-1}. \tag*{(30)} T − 1 = ( I + D − 1 J ) − 1 D − 1 = j = 0 ∑ r comp − 1 ( − D − 1 J ) j D − 1 . (30)
The j j j th term has norm at most ( C N ) j τ f a c − j − 1 (C\sqrt{N})^j\tau_{\mathrm{fac}}^{-j-1} ( C N ) j τ fac − j − 1 . Using τ f a c ≤ 1 \tau_{\mathrm{fac}}\leq1 τ fac ≤ 1 and N ≥ 1 N\geq1 N ≥ 1 to sum these terms gives
∥ T − 1 ∥ ≤ C N ( r c o m p − 1 ) / 2 τ f a c − r c o m p . \lVert T^{-1}\rVert\leq CN^{(r_{\mathrm{comp}}-1)/2}\tau_{\mathrm{fac}}^{-r_{\mathrm{comp}}}. ∥ T − 1 ∥ ≤ C N ( r comp − 1 ) /2 τ fac − r comp .
For
τ f a c = ( 2 C N ( r c o m p − 1 ) / 2 t ) 1 / r c o m p ≤ 1 , \tau_{\mathrm{fac}}=\left(2CN^{(r_{\mathrm{comp}}-1)/2}t\right)^{1/r_{\mathrm{comp}}}\leq1, τ fac = ( 2 C N ( r comp − 1 ) /2 t ) 1/ r comp ≤ 1 ,
this upper bound is ( 2 t ) − 1 (2t)^{-1} ( 2 t ) − 1 . Hence
{ s min ( T ) ≤ t , Ω N } ⊂ ⋃ a = 1 r c o m p { s min ( D a ) < τ f a c , Ω N } . \{s_{\min}(T)\leq t,\ \Omega_N\}\subset\bigcup_{a=1}^{r_{\mathrm{comp}}}\{s_{\min}(D_a)<\tau_{\mathrm{fac}},\ \Omega_N\}. { s m i n ( T ) ≤ t , Ω N } ⊂ a = 1 ⋃ r comp { s m i n ( D a ) < τ fac , Ω N } .
Identity factors do not contribute to this union. Proposition 4.4, applied to the fixed nonzero factors, bounds the right-hand side by
C N K τ f a c ω + C e − c N ≤ C N K + ω ( r c o m p − 1 ) / ( 2 r c o m p ) t ω / r c o m p + C e − c N . CNK\tau_{\mathrm{fac}}^\omega+Ce^{-cN}\leq CNK^{+\omega(r_{\mathrm{comp}}-1)/(2r_{\mathrm{comp}})}t^{\omega/r_{\mathrm{comp}}}+Ce^{-cN}. C N K τ fac ω + C e − c N ≤ C N K + ω ( r comp − 1 ) / ( 2 r comp ) t ω / r comp + C e − c N .
For Gaussian laws we use ω = 1 \omega=1 ω = 1 . If τ f a c \tau_{\mathrm{fac}} τ fac exceeds one, increasing the polynomial prefactor makes the bound at least one. Adding P ( Ω N c ) = q N \mathbb{P}(\Omega_N^c)=q_N P ( Ω N c ) = q N gives the unconditional estimate. If all coefficients vanish, T = I T=I T = I and the estimate is immediate.
The localization error. Since N − 1 ∥ X ℓ ∥ H S 2 = N − 2 ∑ i j ∣ ξ ℓ , i j ∣ 2 N^{-1}\lVert X_\ell\rVert_{\mathrm{HS}}^2=N^{-2}\sum_{ij}|\xi_{\ell,ij}|^2 N − 1 ∥ X ℓ ∥ HS 2 = N − 2 ∑ ij ∣ ξ ℓ , ij ∣ 2 , the weak law gives q N → 0 q_N\to0 q N → 0 . If the law has a ( 2 + ϵ ) (2+\epsilon) ( 2 + ϵ ) moment, set r = 1 + min ( ϵ , 2 ) / 2 ∈ ( 1 , 2 ] r=1+\min(\epsilon,2)/2\in(1,2] r = 1 + min ( ϵ , 2 ) /2 ∈ ( 1 , 2 ] and write Z j = ∣ ξ j ∣ 2 − 1 Z_j=|\xi_j|^2-1 Z j = ∣ ξ j ∣ 2 − 1 , 1 ≤ j ≤ N 2 1\leq j\leq N^2 1 ≤ j ≤ N 2 . Let Z j ′ Z'_j Z j ′ be independent copies and ε j \varepsilon_j ε j independent Rademacher signs. Conditional Jensen’s inequality, symmetry of Z j − Z j ′ Z_j-Z'_j Z j − Z j ′ , and the Rademacher second moment give
E ∣ ∑ j Z j ∣ r ≤ E ∣ ∑ j ( Z j − Z j ′ ) ∣ r = E E ε ∣ ∑ j ε j ( Z j − Z j ′ ) ∣ r ≤ E ( ∑ j ∣ Z j − Z j ′ ∣ 2 ) r / 2 ≤ ∑ j E ∣ Z j − Z j ′ ∣ r ≤ C N 2 E ∣ ∣ ξ ∣ 2 − 1 ∣ r . \begin{aligned}
\mathbb{E}\left|\sum_j Z_j\right|^r\leq\mathbb{E}\left|\sum_j(Z_j-Z'_j)\right|^r=\mathbb{E}\mathbb{E}_{\varepsilon}\left|\sum_j\varepsilon_j(Z_j-Z'_j)\right|^r \\
&\leq\mathbb{E}\left(\sum_j|Z_j-Z'_j|^2\right)^{r/2}\leq\sum_j\mathbb{E}|Z_j-Z'_j|^r\leq CN^{2}\mathbb{E}\left||\xi|^2-1\right|^r.
\end{aligned} E j ∑ Z j r ≤ E j ∑ ( Z j − Z j ′ ) r = E E ε j ∑ ε j ( Z j − Z j ′ ) r ≤ E ( j ∑ ∣ Z j − Z j ′ ∣ 2 ) r /2 ≤ j ∑ E ∣ Z j − Z j ′ ∣ r ≤ C N 2 E ∣ ξ ∣ 2 − 1 r .
Markov’s inequality at threshold N 2 N^2 N 2 yields q N ≤ C N − 2 ( r − 1 ) q_N\leq CN^{-2(r-1)} q N ≤ C N − 2 ( r − 1 ) . For either Gaussian class, ∣ ξ ∣ 2 |\xi|^2 ∣ ξ ∣ 2 has a finite exponential moment in a neighborhood of zero. Since E ∣ ξ ∣ 2 = 1 \mathbb{E}|\xi|^2=1 E ∣ ξ ∣ 2 = 1 , some fixed λ > 0 \lambda>0 λ > 0 satisfies log E e λ ∣ ξ ∣ 2 < 2 λ \log\mathbb{E}e^{\lambda|\xi|^2}<2\lambda log E e λ ∣ ξ ∣ 2 < 2 λ . The exponential Markov inequality at threshold 2 N 2 2N^2 2 N 2 , and a union over ℓ \ell ℓ , give q N ≤ C e − c N 2 q_N\leq Ce^{-cN^2} q N ≤ C e − c N 2 . Combining this with the localized Gaussian estimate completes the proof. □
Least singular values for Wigner tensor pencils
We prove Theorem 1.5 and its consequence for polynomial shifts by matching inverse traces with determinant derivatives in unordered-edge coordinates. We adapt the profile argument of § 3 to unordered edges: an off-diagonal atom enters both endpoint equations, so we leave only edges from an anchor set to its complement unexposed and discard the anchor outputs. The determinant derivative also changes both endpoint entries. We match it to paired inverse traces, then apply relative anti-concentration and triangular reconstruction. The resulting cutoff will enter the Wigner lower logarithmic tail estimate.
In this section abbreviate T N W T_N^{\mathrm{W}} T N W to T N T_N T N and write
T N = I m N + ∑ ℓ = 1 n C ℓ ⊗ H ℓ ( N ) . T_N=I_{mN}+\sum_{\ell=1}^{n}C_\ell\otimes H_\ell^{(N)}. T N = I m N + ℓ = 1 ∑ n C ℓ ⊗ H ℓ ( N ) .
The off-diagonal atoms are independent across ℓ \ell ℓ and unordered edges. For each ℓ \ell ℓ , they have a fixed centered law of absolute variance one, with bounded real or joint planar density. Both types of law may occur in the same pencil. For each ℓ \ell ℓ , the diagonal atoms are independent copies of a fixed centered real law with bounded density and finite second moment. The two coordinates of a planar atom may be dependent.
We use the following real coordinates. For i < k i<k i < k , write ζ ℓ , i k = x ℓ , { i , k } + i y ℓ , { i , k } \zeta_{\ell,ik}=x_{\ell,\{i,k\}}+iy_{\ell,\{i,k\}} ζ ℓ , ik = x ℓ , { i , k } + i y ℓ , { i , k } , omitting y y y for real entry laws, and set d ℓ , i = N ( H ℓ ( N ) ) i i d_{\ell,i}=\sqrt{N}(H_{\ell}^{(N)})_{ii} d ℓ , i = N ( H ℓ ( N ) ) ii . We use ε i k = 1 \varepsilon_{ik}=1 ε ik = 1 for i < k i<k i < k and ε i k = − 1 \varepsilon_{ik}=-1 ε ik = − 1 for i > k i>k i > k . Thus, for i ≠ k i\ne k i = k ,
( H ℓ ( N ) ) i k = N − 1 / 2 ( x ℓ , { i , k } + i ε i k y ℓ , { i , k } ) . (H_{\ell}^{(N)})_{ik}=N^{-1/2}(x_{\ell,\{i,k\}}+i\varepsilon_{ik}y_{\ell,\{i,k\}}). ( H ℓ ( N ) ) ik = N − 1/2 ( x ℓ , { i , k } + i ε ik y ℓ , { i , k } ) .
Let Y ℓ , x Y_{\ell,x} Y ℓ , x be the real symmetric matrix with off-diagonal entries x ℓ , { i , k } / N x_{\ell,\{i,k\}}/\sqrt{N} x ℓ , { i , k } / N and diagonal d ℓ , i / N d_{\ell,i}/\sqrt{N} d ℓ , i / N , and let Y ℓ , y Y_{\ell,y} Y ℓ , y be the real skew-symmetric matrix with entries ε i k y ℓ , { i , k } / N \varepsilon_{ik}y_{\ell,\{i,k\}}/\sqrt{N} ε ik y ℓ , { i , k } / N and zero diagonal. Then H ℓ ( N ) = Y ℓ , x + i Y ℓ , y H_{\ell}^{(N)}=Y_{\ell,x}+iY_{\ell,y} H ℓ ( N ) = Y ℓ , x + i Y ℓ , y and
∥ Y ℓ , x ∥ H S 2 + ∥ Y ℓ , y ∥ H S 2 = ∥ H ℓ ( N ) ∥ H S 2 . (31) \lVert Y_{\ell,x}\rVert_{\mathrm{HS}}^{2}+\lVert Y_{\ell,y}\rVert_{\mathrm{HS}}^{2}=\lVert H_{\ell}^{(N)}\rVert_{\mathrm{HS}}^{2}. \tag*{(31)} ∥ Y ℓ , x ∥ HS 2 + ∥ Y ℓ , y ∥ HS 2 = ∥ H ℓ ( N ) ∥ HS 2 . (31)
Throughout this section, y y y terms are omitted for real entry laws. In sums indexed by h h h , the value h = x h=x h = x is always included, and h = y h=y h = y is included only for planar entry laws. We write Ω N W \Omega_{N}^{\mathrm W} Ω N W and q N W q_{N}^{\mathrm W} q N W for the localization event and its complementary probability in (5) . We use (31) to bound both the pencil norm and the contribution of small singular directions of the profile coefficient map.
A profile estimate with unordered edges
The dimension surplus depends only on the coefficient maps. To apply the uniform covering and projection arguments here, we prove the probability estimate for whole unordered edges and include the diagonal contribution in the approximation error.
Let V V V be a real Hilbert space of fixed dimension p ≥ 1 p\ge1 p ≥ 1 , let D ⊂ V D\subset V D ⊂ V be a fixed real subspace, and let L ℓ , x , L ℓ , y : V → V L_{\ell,x},L_{\ell,y}:V\to V L ℓ , x , L ℓ , y : V → V be fixed real linear maps, with L ℓ , y L_{\ell,y} L ℓ , y present only for planar entry laws. Assume that no nonzero real subspace of D D D is invariant under all the available maps. Define, for a ∈ [ N ] a\in[N] a ∈ [ N ] and v ∈ D N v\in D^{N} v ∈ D N ,
( D a W v ) k = v k + 1 N ∑ i ≠ k , ℓ ( x ℓ , { i , k } L ℓ , x v i + ε i k y ℓ , { i , k } L ℓ , y v i ) + 1 N ∑ ℓ d ℓ , k L ℓ , x v k , k ≠ a . (32) (\mathcal{D}_{a}^{\mathrm W}v)_{k}=v_{k}+\frac{1}{\sqrt{N}}\sum_{i\ne k,\ell}\left(x_{\ell,\{i,k\}}L_{\ell,x}v_{i}+\varepsilon_{ik}y_{\ell,\{i,k\}}L_{\ell,y}v_{i}\right)+\frac{1}{\sqrt{N}}\sum_{\ell}d_{\ell,k}L_{\ell,x}v_{k},\qquad k\ne a. \tag*{(32)} ( D a W v ) k = v k + N 1 i = k , ℓ ∑ ( x ℓ , { i , k } L ℓ , x v i + ε ik y ℓ , { i , k } L ℓ , y v i ) + N 1 ℓ ∑ d ℓ , k L ℓ , x v k , k = a . (32)
The same formula on V N V^{N} V N defines D ~ a W \widetilde{\mathcal{D}}_{a}^{\mathrm W} D a W . Equation (31) implies, on Ω N W \Omega_{N}^{\mathrm W} Ω N W , that
∥ D ~ a W ∥ ≤ C 0 N \lVert\widetilde{\mathcal{D}}_{a}^{\mathrm W}\rVert\le C_{0}\sqrt{N} ∥ D a W ∥ ≤ C 0 N
simultaneously for all a a a .
The same localization gives a bound on the coordinate squares:
∑ ℓ ( 2 ∑ i < k ( x ℓ , { i , k } 2 + y ℓ , { i , k } 2 ) + ∑ i d ℓ , i 2 ) ≤ 2 n N 2 . \sum_{\ell}\left(2\sum_{i<k}\left(x_{\ell,\{i,k\}}^{2}+y_{\ell,\{i,k\}}^{2}\right)+\sum_{i}d_{\ell,i}^{2}\right)\le2nN^{2}. ℓ ∑ ( 2 i < k ∑ ( x ℓ , { i , k } 2 + y ℓ , { i , k } 2 ) + i ∑ d ℓ , i 2 ) ≤ 2 n N 2 .
This also bounds the inactive part of the profile map, including the diagonal terms. Constants below depend on the fixed maps and laws.
Lemma 6.1 (Unordered-edge profile estimate). If D ≠ 0 \mathcal{D} \ne0 D = 0 , there are E ∗ < ∞ E_* < \infty E ∗ < ∞ and N 0 N_0 N 0 such that, for N ≥ N 0 N \ge N_0 N ≥ N 0 ,
P ( Ω N W ∩ { ∃ a : s min ( D a W ) ≤ N − E ∗ } ) ≤ C e − c N . \mathbb{P}\left(\Omega_N^W \cap\{\exists a : s_{\min}(\mathcal{D}_a^W) \le N^{-E_*}\}\right) \le C e^{-cN}. P ( Ω N W ∩ { ∃ a : s m i n ( D a W ) ≤ N − E ∗ } ) ≤ C e − c N . The conclusion uses no moment above two and no independence within a planar off-diagonal atom.
Proof. For a nonzero E ⊂ D E \subset\mathcal{D} E ⊂ D , let r r r be the number of available maps L ℓ , x , L ℓ , y L_{\ell,x}, L_{\ell,y} L ℓ , x , L ℓ , y , and let Ψ E : E r → V \Psi_E : E^r \to V Ψ E : E r → V sum them. Define its active image Y σ ( E ) Y_\sigma(E) Y σ ( E ) and the right singular space C τ , σ ( E ) C_{\tau,\sigma}(E) C τ , σ ( E ) of P Y σ ( E ) ⊥ ∣ E P_{Y_\sigma(E)^\perp}|_E P Y σ ( E ) ⊥ ∣ E as in § 3 . Lemma 3.2 , applied to this list of maps, gives fixed σ 0 , τ > 0 \sigma_0,\tau> 0 σ 0 , τ > 0 with
y : = dim Y σ ( E ) ≥ c E + 1 , c E : = dim C τ , σ ( E ) , 0 < σ ≤ σ 0 . (33) y := \dim Y_\sigma(E) \ge c_E + 1,\qquad c_E := \dim C_{\tau,\sigma}(E),\qquad0 < \sigma\le\sigma_0. \tag*{(33)} y := dim Y σ ( E ) ≥ c E + 1 , c E := dim C τ , σ ( E ) , 0 < σ ≤ σ 0 . (33)
All dimensions here are real.
Conditioning on whole edges. Fix a deterministic v ∈ E N v \in E^N v ∈ E N with ∥ v ∥ 2 ≤ C 1 \lVert v\rVert^2 \le C_1 ∥ v ∥ 2 ≤ C 1 and Γ v ⪰ γ P E \Gamma v \succeq\gamma P_E Γ v ⪰ γ P E , where d = dim E d = \dim E d = dim E . Lemma 3.4 selects, before any randomness is exposed, an anchor set I I I with ∣ I ∣ = d |I| = d ∣ I ∣ = d and
Γ v , I : = ∑ i ∈ I v i ⊗ R v i ⪰ c d , C 1 γ d N − d P E . \Gamma_{v,I} := \sum_{i\in I} v_i \otimes_{\mathbb{R}} v_i \succeq c_{d,C_1}\gamma^d N^{-d}P_E. Γ v , I := i ∈ I ∑ v i ⊗ R v i ⪰ c d , C 1 γ d N − d P E .
Let J = [ N ] ∖ ( I ∪ { a } ) J = [N] \setminus(I \cup\{a\}) J = [ N ] ∖ ( I ∪ { a }) and expose every atom except the whole edge blocks on { i , k } \{i,k\} { i , k } with i ∈ I i \in I i ∈ I and k ∈ J k \in J k ∈ J . In particular, all diagonal atoms are exposed. The remaining random part of output k ∈ J k \in J k ∈ J is
Z k = N − 1 / 2 ∑ i ∈ I , ℓ ( x ℓ , { i , k } L ℓ , x v i + ε i k y ℓ , { i , k } L ℓ , y v i ) . Z_k = N^{-1/2}\sum_{i\in I,\ell}\left(x_{\ell,\{i,k\}}L_{\ell,x}v_i+\varepsilon_{ik}y_{\ell,\{i,k\}}L_{\ell,y}v_i\right). Z k = N − 1/2 i ∈ I , ℓ ∑ ( x ℓ , { i , k } L ℓ , x v i + ε ik y ℓ , { i , k } L ℓ , y v i ) .
The second occurrence of every unexposed block is in an anchor output, which has been discarded. The remaining outputs therefore use disjoint independent blocks. All other terms of their equations are fixed translations.
For Y = Y σ ( E ) Y = Y_\sigma(E) Y = Y σ ( E ) , the real linear map from the unexposed coordinates to P Y Z k P_Y Z_k P Y Z k has Euclidean Gram operator
1 N P Y ∑ ℓ ( L ℓ , x Γ v , I L ℓ , x ∗ + L ℓ , y Γ v , I L ℓ , y ∗ ) P Y ⪰ c σ 2 γ d N − d − 1 I Y . \frac{1}{N}P_Y\sum_\ell\left(L_{\ell,x}\Gamma_{v,I}L_{\ell,x}^*+L_{\ell,y}\Gamma_{v,I}L_{\ell,y}^*\right)P_Y \succeq c\sigma^2\gamma^dN^{-d-1}I_Y. N 1 P Y ℓ ∑ ( L ℓ , x Γ v , I L ℓ , x ∗ + L ℓ , y Γ v , I L ℓ , y ∗ ) P Y ⪰ c σ 2 γ d N − d − 1 I Y .
Each y y y coefficient is multiplied by ε i k \varepsilon_{ik} ε ik , whose square is one, so its outer product is the same at either endpoint. The lower bound follows from the anchor estimate and Ψ E Ψ E ∗ ⪰ σ 2 I Y \Psi_E\Psi_E^* \succeq\sigma^2 I_Y Ψ E Ψ E ∗ ⪰ σ 2 I Y on Y Y Y . This calculation concerns the deterministic coordinate map, so dependence between the coordinates of a planar block does not affect it.
Fix κ = 1 / ( 8 p ) \kappa= 1/(8p) κ = 1/ ( 8 p ) . Since only finitely many laws occur, we can choose R 0 R_0 R 0 so large that, for every anchor set of size at most p p p , all its n ∣ I ∣ n|I| n ∣ I ∣ edge blocks in one output have norm at most R 0 R_0 R 0 with probability at least 1 − κ / 4 1-\kappa/4 1 − κ /4 . This choice uses only that each law is a probability measure. For a fixed anchor set, the good-output indicators involve disjoint edges and are independent. Chernoff’s inequality, followed by a union over at most p N p pN^p p N p anchor sets, gives an event A N W \mathcal{A}_N^W A N W with P ( ( A N W ) c ) ≤ C e − c N \mathbb{P}((\mathcal{A}_N^W)^c) \le Ce^{-cN} P (( A N W ) c ) ≤ C e − c N , on which at most κ N \kappa N κ N outputs fail the cutoff for any anchor set.
Condition on the pattern of outputs satisfying the cutoff. In each such output the unexposed edge blocks are independent copies of their original laws conditioned to lie in the radius-R 0 R_0 R 0 ball. Each planar block is conditioned as a whole. At most n p np n p blocks occur, each of real dimension one or two, so their product density and support radius are bounded by fixed constants. The preceding Gram bound and Lemma 3.5 then give the one-output small-ball factor
C σ − y γ − d y / 2 N ( d + 1 ) y / 2 δ y . C\sigma^{-y}\gamma^{-dy/2}N^{(d+1)y/2}\delta^{y}. C σ − y γ − d y /2 N ( d + 1 ) y /2 δ y .
On A N W \mathcal{A}^{\mathrm{W}}_{N} A N W the chosen anchor set has at least b N ( d ) = ⌊ ( 1 − κ ) N ⌋ − d − 1 b_{N}(d)=\lfloor(1-\kappa)N\rfloor-d-1 b N ( d ) = ⌊( 1 − κ ) N ⌋ − d − 1 good outputs in J J J . Replacing the global cutoff event by this assertion for the chosen I I I , we multiply the one-output factors and average over the patterns. Their probabilities sum to at most one, so we obtain
P { ∥ D a W v ∥ 2 ≤ δ , A N W } ≤ min { 1 , ( C σ − y γ − d y / 2 N ( d + 1 ) y / 2 δ y ) b N ( d ) } . (34) \mathbb{P}\{\lVert\mathcal{D}^{\mathrm{W}}_{a}v\rVert_{2}\leq\delta,\mathcal{A}^{\mathrm{W}}_{N}\}\leq\min\left\{1,\left(C\sigma^{-y}\gamma^{-dy/2}N^{(d+1)y/2}\delta^{y}\right)^{b_{N}(d)}\right\}. \tag*{(34)} P {∥ D a W v ∥ 2 ≤ δ , A N W } ≤ min { 1 , ( C σ − y γ − d y /2 N ( d + 1 ) y /2 δ y ) b N ( d ) } . (34)
Integrating over the first exposure proves the estimate for the intersection with A N W \mathcal{A}^{\mathrm{W}}_{N} A N W without conditioning on that event; no restriction by Ω N W \Omega^{\mathrm{W}}_{N} Ω N W was used.
Uniformity over profiles. The fixed-profile bound has the same exponent and scale factors as in the independent model. To apply the subspace and tube nets of Lemma 3.7, we bound the remaining contribution in the inactive singular directions. At a net center E E E in the family with a spectral gap, the random sum in output k k k is Ψ E t k \Psi_{E}t_{k} Ψ E t k , where
t ℓ , x , k = N − 1 / 2 ( ∑ i ≠ k x ℓ , { i , k } v i + d ℓ , k v k ) , t_{\ell,x,k}=N^{-1/2}\left(\sum_{i\ne k}x_{\ell,\{i,k\}}v_{i}+d_{\ell,k}v_{k}\right), t ℓ , x , k = N − 1/2 i = k ∑ x ℓ , { i , k } v i + d ℓ , k v k ,
t ℓ , y , k = N − 1 / 2 ∑ i ≠ k ε i k y ℓ , { i , k } v i . t_{\ell,y,k}=N^{-1/2}\sum_{i\ne k}\varepsilon_{ik}y_{\ell,\{i,k\}}v_{i}. t ℓ , y , k = N − 1/2 i = k ∑ ε ik y ℓ , { i , k } v i .
Cauchy–Schwarz and the preceding bound on the coordinate squares, including the diagonal terms, give
∥ t k ∥ 2 ≤ ∥ v ∥ 2 2 N ∑ ℓ ( ∑ i ≠ k x ℓ , { i , k } 2 + y ℓ , { i , k } 2 ) + d ℓ , k 2 \lVert t_{k}\rVert^{2}\leq\frac{\lVert v\rVert_{2}^{2}}{N}\sum_{\ell}\left(\sum_{i\ne k}x_{\ell,\{i,k\}}^{2}+y_{\ell,\{i,k\}}^{2}\right)+d_{\ell,k}^{2} ∥ t k ∥ 2 ≤ N ∥ v ∥ 2 2 ℓ ∑ i = k ∑ x ℓ , { i , k } 2 + y ℓ , { i , k } 2 + d ℓ , k 2
≤ 2 n N ∥ v ∥ 2 2 . \leq2nN\lVert v\rVert_{2}^{2}. ≤ 2 n N ∥ v ∥ 2 2 .
Thus projection of the random sum onto the orthogonal complement of the active image leaves a contribution of at most C N η C\sqrt{N}\eta C N η . When this is smaller than the small-image radius, the profile coordinates are confined to the same tubes as before. The operator bound (6.3) controls rounding errors, while (33) gives the excess of active output dimensions over the dimensions of the profile net. The same Grassmannian and tube construction therefore applies, including its Gram stability estimate. Summing the integrated probabilities (34) over this net leaves the dimension surplus y b N ( d ) − c E ( N − 1 ) − d ≥ 7 N / 8 − O p ( 1 ) yb_{N}(d)-c_{E}(N-1)-d\geq7N/8-O_{p}(1) y b N ( d ) − c E ( N − 1 ) − d ≥ 7 N /8 − O p ( 1 ) , yielding
C ∗ N N K 0 N γ − p 2 N σ − p N δ N / 2 − L 0 (35) C_{*}^{N}N^{K_{0}N}\gamma^{-p_{2}N}\sigma^{-pN}\delta^{N/2-L_{0}} \tag*{(35)} C ∗ N N K 0 N γ − p 2 N σ − pN δ N /2 − L 0 (35)
for the probability of a small image among regular profiles in Ω N W ∩ A N W \Omega^{\mathrm{W}}_{N}\cap\mathcal{A}^{\mathrm{W}}_{N} Ω N W ∩ A N W , under the same conditions 0 < δ ≤ c ∗ γ 0<\delta\leq c_{*}\gamma 0 < δ ≤ c ∗ γ and C ∗ N η ≤ δ C_{*}\sqrt{N}\eta\leq\delta C ∗ N η ≤ δ . The constants depend on the fixed maps and entry laws.
Spectral scales and projection of profiles. The remaining reductions use only the uniform probability bound and the localized operator norm. Applying the spectral scales of Lemma 3.8 to (35) and the surplus gives a cutoff N − e ( D ) N^{-e(D)} N − e ( D ) for Γ v ⪰ N − D P E \Gamma_{v}\succeq N^{-D}P_{E} Γ v ⪰ N − D P E , with failure probability p e − 2 N log N p e^{-2N\log N} p e − 2 N l o g N . We then project away the eigenvectors corresponding to small Gram eigenvalues as in the proof of Theorem 3.1, using the C 0 N C_{0}\sqrt{N} C 0 N operator bound. With the thresholds D 1 = 1 D_{1}=1 D 1 = 1 , D d + 1 ≥ max { D d + 1 , 2 e ( D d ) + 4 } D_{d+1}\geq\max\{D_{d}+1,2e(D_{d})+4\} D d + 1 ≥ max { D d + 1 , 2 e ( D d ) + 4 } , the projection error in the image is at most 1 2 N − e ( D d ) \frac{1}{2}N^{-e(D_{d})} 2 1 N − e ( D d ) at the stopping dimension. Hence E ∗ > max d e ( D d ) + 1 E_{*}>\max_{d}e(D_{d})+1 E ∗ > max d e ( D d ) + 1 suffices. Taking a union over the finitely many dimensions and the N N N choices of a a a , and adding P ( ( A N W ) c ) \mathbb{P}((\mathcal{A}_N^W)^c) P (( A N W ) c ) proves the lemma. □ \square □
Mixed real and planar relative anti-concentration
The inverse estimate uses the real gradient over independent atom blocks. Off-diagonal atoms may be planar, while diagonal atoms are real, so the relative estimate must allow both types. For a planar block, we bound the area of the small-value set before applying its joint density, without conditioning one coordinate on the other.
Lemma 6.2. Let F F F be a nonzero complex-valued polynomial of degree at most d d d in each real coordinate. Its independent atom blocks comprise k 1 k_1 k 1 real variables with densities bounded by M M M , and k 2 k_2 k 2 planar variables with joint densities bounded by M M M and second moments at most M 2 M_2 M 2 . Write K c o o r d = k 1 + 2 k 2 ≥ 1 K_{\mathrm{coord}} = k_1 + 2k_2 \ge1 K coord = k 1 + 2 k 2 ≥ 1 . Then
P { ∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C d , M , M 2 K c o o r d 3 / 2 t 1 / 2 , 0 < t ≤ 1. \mathbb{P}\{|F| \le t\|\nabla F\|_2\} \le C_{d,M,M_2} K_{\mathrm{coord}}^{3/2} t^{1/2}, \qquad0 < t \le1. P { ∣ F ∣ ≤ t ∥∇ F ∥ 2 } ≤ C d , M , M 2 K coord 3/2 t 1/2 , 0 < t ≤ 1. If k 2 = 0 k_2 = 0 k 2 = 0 , one can replace t 1 / 2 t^{1/2} t 1/2 by t t t .
Proof. Outside the null zero set of F F F , the event in question implies, for some real coordinate j j j ,
∣ F ∣ ≤ t K c o o r d ∣ ∂ j F ∣ . |F| \le t\sqrt{K_{\mathrm{coord}}}\,|\partial_j F|. ∣ F ∣ ≤ t K coord ∣ ∂ j F ∣. For a real atom, fix the other whole blocks and use the root-factorization estimate from Lemma 4.3 . The conditional probability is at most 2 M d 2 t K c o o r d 2Md^2t\sqrt{K_{\mathrm{coord}}} 2 M d 2 t K coord . For a coordinate in a planar atom, fix only the other blocks. Within the radius-R c u t R_{\mathrm{cut}} R cut disk, the one-variable root estimate bounds every slice by length 2 d 2 t K c o o r d 2d^2t\sqrt{K_{\mathrm{coord}}} 2 d 2 t K coord . Fubini therefore bounds the admissible area by 4 d 2 R c u t t K c o o r d 4d^2R_{\mathrm{cut}}t\sqrt{K_{\mathrm{coord}}} 4 d 2 R cut t K coord . Multiplying by the joint density bound gives the conditional probability estimate without assuming independence of the two coordinates. Summing over the coordinates and using the second-moment bound outside the planar disks gives
P { ∣ F ∣ ≤ t ∥ ∇ F ∥ 2 } ≤ C M d 2 K c o o r d 3 / 2 ( t + R c u t t ) + k 2 M 2 R c u t − 2 . \mathbb{P}\{|F| \le t\|\nabla F\|_2\} \le CMd^2K_{\mathrm{coord}}^{3/2}(t+R_{\mathrm{cut}}t)+k_2M_2R_{\mathrm{cut}}^{-2}. P { ∣ F ∣ ≤ t ∥∇ F ∥ 2 } ≤ C M d 2 K coord 3/2 ( t + R cut t ) + k 2 M 2 R cut − 2 . Taking R c u t = t − 1 / 2 R_{\mathrm{cut}}=t^{-1/2} R cut = t − 1/2 proves the square-root bound. The exceptional coefficient and slice zero sets are null under the product of the absolutely continuous block laws, as in Lemma 4.3 . When there are no planar atoms, the real-variable estimate alone gives the stated linear bound. □ \square □
The doubled algebra and the determinant gradient
We choose the doubled generators so that their traces encode the symmetric and skew-symmetric variations of an unordered edge. The profile estimate gives a trace frame on the adjoint algebra. We identify each trace with an edge or diagonal derivative and sum the frame inequalities over the rows.
Let ( C ℓ ) (C_\ell) ( C ℓ ) be a nonzero complex irreducible coefficient family. Since an edge derivative contains the sum or difference of the traces at its endpoints, define
D ℓ , x = diag ( C ℓ , C ℓ T ) , D ℓ , y = diag ( i C ℓ , − i C ℓ T ) , B = alg C ( I , D ℓ , x , D ℓ , y ) , (36) D_{\ell,x}=\operatorname{diag}(C_\ell,C_\ell^{\mathsf T}), \qquad D_{\ell,y}=\operatorname{diag}(iC_\ell,-iC_\ell^{\mathsf T}), \qquad\mathcal{B}=\operatorname{alg}_{\mathbb{C}}(I,D_{\ell,x},D_{\ell,y}), \tag*{(36)} D ℓ , x = diag ( C ℓ , C ℓ T ) , D ℓ , y = diag ( i C ℓ , − i C ℓ T ) , B = alg C ( I , D ℓ , x , D ℓ , y ) , (36)
again omitting y y y for real entry laws. The associated pencil satisfies
T ^ N = I + ∑ ℓ ( D ℓ , x ⊗ Y ℓ , x + D ℓ , y ⊗ Y ℓ , y ) ≃ T N ⊕ T N T . (37) \widehat{T}_N=I+\sum_\ell(D_{\ell,x}\otimes Y_{\ell,x}+D_{\ell,y}\otimes Y_{\ell,y})\simeq T_N\oplus T_N^{\mathsf T}. \tag*{(37)} T N = I + ℓ ∑ ( D ℓ , x ⊗ Y ℓ , x + D ℓ , y ⊗ Y ℓ , y ) ≃ T N ⊕ T N T . (37)
The upper block is T N T_N T N ; the lower one is T N T T_N^{\mathsf T} T N T because Y ℓ , x T = Y ℓ , x Y_{\ell,x}^{\mathsf T}=Y_{\ell,x} Y ℓ , x T = Y ℓ , x and Y ℓ , y T = − Y ℓ , y Y_{\ell,y}^{\mathsf T}=-Y_{\ell,y} Y ℓ , y T = − Y ℓ , y .
The two coordinate representations of B \mathcal{B} B are generated by C ℓ C_\ell C ℓ and C ℓ T C_\ell^{\mathsf T} C ℓ T , up to scalar multiples. Both are irreducible, the second by the annihilator argument in Lemma 2.3 . The same lemma gives a nondegenerate trace form and zero common kernel for these doubled generators. On the real Hilbert space V = B ∗ V=\mathcal{B}^{*} V = B ∗ define
O ( Φ ) = ( Tr ( Φ ∗ D ℓ , h ) ) ℓ , h , D = ker O . \mathcal{O}(\Phi)=\left(\operatorname{Tr}(\Phi^{*}D_{\ell,h})\right)_{\ell,h},\qquad\mathcal{D}=\ker\mathcal{O}. O ( Φ ) = ( Tr ( Φ ∗ D ℓ , h ) ) ℓ , h , D = ker O .
Left multiplication preserves V V V , since D ℓ , h ∗ b ∗ = ( b D ℓ , h ) ∗ D_{\ell,h}^{*}b^{*}=(bD_{\ell,h})^{*} D ℓ , h ∗ b ∗ = ( b D ℓ , h ) ∗ . The right-ideal argument of Lemma 2.5 therefore applies and gives (16) for this D \mathcal{D} D .
If R = T N − 1 R=T_{N}^{-1} R = T N − 1 , the spatial blocks of the doubled inverse are
R ^ i j = diag ( R i j , R j i T ) ∈ B . \widehat{R}_{ij}=\operatorname{diag}(R_{ij},R_{ji}^{\mathsf{T}})\in\mathcal{B}. R ij = diag ( R ij , R j i T ) ∈ B .
Membership follows from the Cayley–Hamilton argument of Lemma 2.4 , which requires only that the spatial blocks of the doubled pencil belong to the unital algebra. Taking adjoints in R ^ T ^ N = I \widehat{R}\widehat{T}_{N}=I R T N = I shows that Φ i = R ^ a i ∗ \Phi_i=\widehat{R}_{ai}^{*} Φ i = R ai ∗ solves (32) with zero outputs for k ≠ a k\ne a k = a , with L ℓ , h Φ = D ℓ , h ∗ Φ L_{\ell,h}\Phi=D_{\ell,h}^{*}\Phi L ℓ , h Φ = D ℓ , h ∗ Φ . For this profile, projection onto the kernel of the trace map and its orthogonal complement gives the same decomposition as in Proposition 4.1 . The part in the kernel is controlled by Lemma 6.1 ; the contribution of its complement to the equation is bounded by (6.3). Injectivity of the trace map on that fixed complement then gives a finite β \beta β and an event E N W ⊂ Ω N W \mathcal{E}_{N}^{\mathsf{W}}\subset\Omega_{N}^{\mathsf{W}} E N W ⊂ Ω N W , with P ( Ω N W ∖ E N W ) ≤ C e − c N \mathbb{P}(\Omega_{N}^{\mathsf{W}}\setminus\mathcal{E}_{N}^{\mathsf{W}})\le Ce^{-cN} P ( Ω N W ∖ E N W ) ≤ C e − c N , on which every inverse-row profile satisfies
1 N ∑ i , ℓ , h ∣ Tr ( Φ i ∗ D ℓ , h ) ∣ 2 ≥ N − β ∑ i ∥ Φ i ∥ H S 2 . (38) \frac{1}{N}\sum_{i,\ell,h}\left|\operatorname{Tr}(\Phi_i^{*}D_{\ell,h})\right|^{2}\ge N^{-\beta}\sum_i\|\Phi_i\|_{\mathrm{HS}}^{2}. \tag*{(38)} N 1 i , ℓ , h ∑ ∣ Tr ( Φ i ∗ D ℓ , h ) ∣ 2 ≥ N − β i ∑ ∥ Φ i ∥ HS 2 . (38)
If D = 0 \mathcal{D}=0 D = 0 , finite-dimensional injectivity alone gives this estimate, with no use of the dual lemma.
The trace observations control each doubled inverse row, so summing their squared moduli controls the full Hilbert–Schmidt inverse norm. We express them through normalized derivatives of the original determinant; the doubled pencil serves only to pair the endpoint traces. Let F = det T N F=\det T_N F = det T N , viewed as a polynomial in the real coordinates of the independent atom blocks. The gradient contains one derivative for each available coordinate of each unordered edge and each diagonal atom. An x x x coordinate changes the two endpoint entries with the same sign, a y y y coordinate changes them with opposite imaginary signs, and a diagonal coordinate changes only one entry. Jacobi’s formula consequently gives, for i < j i<j i < j ,
∂ x ℓ , { i , j } F F = N − 1 / 2 [ Tr ( R j i C ℓ ) + Tr ( R i j C ℓ ) ] . (39) \frac{\partial_{x_{\ell,\{i,j\}}}F}{F}
=N^{-1/2}\left[\operatorname{Tr}(R_{ji}C_\ell)+\operatorname{Tr}(R_{ij}C_\ell)\right]. \tag*{(39)} F ∂ x ℓ , { i , j } F = N − 1/2 [ Tr ( R j i C ℓ ) + Tr ( R ij C ℓ ) ] . (39)
∂ y ℓ , { i , j } F F = i N − 1 / 2 [ Tr ( R j i C ℓ ) − Tr ( R i j C ℓ ) ] . (40) \frac{\partial_{y_{\ell,\{i,j\}}}F}{F}
=iN^{-1/2}\left[\operatorname{Tr}(R_{ji}C_\ell)-\operatorname{Tr}(R_{ij}C_\ell)\right]. \tag*{(40)} F ∂ y ℓ , { i , j } F = i N − 1/2 [ Tr ( R j i C ℓ ) − Tr ( R ij C ℓ ) ] . (40)
∂ d ℓ , i F F = N − 1 / 2 Tr ( R i i C ℓ ) . (41) \frac{\partial_{d_{\ell,i}}F}{F}=N^{-1/2}\operatorname{Tr}(R_{ii}C_\ell). \tag*{(41)} F ∂ d ℓ , i F = N − 1/2 Tr ( R ii C ℓ ) . (41)
Each coordinate changes T N T_N T N by rank at most 2 m 2m 2 m , so the degree of F F F in that coordinate is at most 2 m 2m 2 m . Since F ( 0 ) = 1 F(0)=1 F ( 0 ) = 1 , its zero set is null under the product of the continuous atom laws.
To sum these observations, write A = Tr ( R a i C ℓ ) A=\operatorname{Tr}(R_{ai}C_\ell) A = Tr ( R ai C ℓ ) and B = Tr ( R i a C ℓ ) B=\operatorname{Tr}(R_{ia}C_\ell) B = Tr ( R ia C ℓ ) . Direct multiplication of the doubled blocks gives
Tr ( R ^ a i D ℓ , x ) = A + B , Tr ( R ^ a i D ℓ , y ) = i ( A − B ) . \operatorname{Tr}(\widehat{R}_{ai}D_{\ell,x})=A+B,\qquad\operatorname{Tr}(\widehat{R}_{ai}D_{\ell,y})=i(A-B). Tr ( R ai D ℓ , x ) = A + B , Tr ( R ai D ℓ , y ) = i ( A − B ) .
For a ≠ i a\ne i a = i , these are N \sqrt{N} N times the corresponding unordered-edge derivatives divided by F F F , up to the endpoint sign of the y y y term. Each edge occurs twice in the sum over ( a , i ) (a,i) ( a , i ) . For a = i a=i a = i , the x x x observation is 2 Tr ( R i i C ℓ ) 2\operatorname{Tr}(R_{ii}C_\ell) 2 Tr ( R ii C ℓ ) and the y y y observation vanishes, whereas the diagonal derivative has only one copy of Tr ( R i i C ℓ ) \operatorname{Tr}(R_{ii}C_\ell) Tr ( R ii C ℓ ) . Write ∇ o f f \nabla_{\mathrm{off}} ∇ off and ∇ d i a g \nabla_{\mathrm{diag}} ∇ diag for the derivatives in the unordered-edge and diagonal coordinates, respectively. Thus the full sum S S S of squared observations satisfies
S = 2 N ∥ ∇ o f f F / F ∥ 2 2 + 4 N ∥ ∇ d i a g F / F ∥ 2 2 ≤ 4 N ∥ ∇ F / F ∥ 2 2 , ∑ a , i ∥ R a i ∥ H S 2 = 2 ∥ R ∥ H S 2 . (42) S = 2N\left\lVert\nabla_{\mathrm{off}}F/F\right\rVert_{2}^{2} + 4N\left\lVert\nabla_{\mathrm{diag}}F/F\right\rVert_{2}^{2} \le4N\left\lVert\nabla F/F\right\rVert_{2}^{2}, \qquad\sum_{a,i}\left\lVert R_{ai}\right\rVert_{\mathrm{HS}}^{2} = 2\left\lVert R\right\rVert_{\mathrm{HS}}^{2}. \tag*{(42)} S = 2 N ∥ ∇ off F / F ∥ 2 2 + 4 N ∥ ∇ diag F / F ∥ 2 2 ≤ 4 N ∥ ∇ F / F ∥ 2 2 , a , i ∑ ∥ R ai ∥ HS 2 = 2 ∥ R ∥ HS 2 . (42)
Sum (38) over a a a and use (42) to obtain, on E N W \mathcal{E}_{N}^{\mathrm{W}} E N W outside the null singular set,
∥ T N − 1 ∥ ≤ ∥ T N − 1 ∥ H S ≤ 2 N β / 2 ∥ ∇ F ∥ 2 / ∣ F ∣ . (43) \left\lVert T_{N}^{-1}\right\rVert\le\left\lVert T_{N}^{-1}\right\rVert_{\mathrm{HS}} \le\sqrt{2}N^{\beta/2}\left\lVert\nabla F\right\rVert_{2}/\lvert F\rvert. \tag*{(43)} T N − 1 ≤ T N − 1 HS ≤ 2 N β /2 ∥ ∇ F ∥ 2 / ∣ F ∣ . (43)
Completion of the Wigner least singular value theorem
Relative anti-concentration now gives the tail on each irreducible factor. For arbitrary coefficients, localization bounds the off-diagonal blocks and the finite inverse expansion combines the factor estimates. The localization error and Gaussian improvement are then estimated as in the independent model.
Proof of Theorem 1.5 . For a nonzero irreducible family, (43) implies that s min ( T N ) ≤ t s_{\min}(T_N) \le t s m i n ( T N ) ≤ t on E N W \mathcal{E}_{N}^{\mathrm{W}} E N W only if
∣ F ∣ ≤ 2 t N β / 2 ∥ ∇ F ∥ 2 . \lvert F\rvert\le\sqrt{2}\,tN^{\beta/2}\left\lVert\nabla F\right\rVert_{2}. ∣ F ∣ ≤ 2 t N β /2 ∥ ∇ F ∥ 2 .
There are O ( N 2 ) O(N^{2}) O ( N 2 ) independent atom blocks. If all off-diagonal entry laws are real, Lemma 4.3 (i) bounds this probability by C N K t CN^{K}t C N K t ; otherwise Lemma 6.2 bounds it by C N K t 1 / 2 CN^{K}t^{1/2} C N K t 1/2 . For relative radii above one the same conclusion follows from the trivial probability bound after increasing the polynomial prefactor. Thus, with ω = 1 \omega= 1 ω = 1 when all entry laws are real and ω = 1 / 2 \omega= 1/2 ω = 1/2 otherwise,
P { s min ( T N ) ≤ t , E N W } ≤ C N K t ω + C e − c N . \mathbb{P}\{s_{\min}(T_N) \le t,\mathcal{E}_{N}^{\mathrm{W}}\} \le CN^{K}t^{\omega} + Ce^{-cN}. P { s m i n ( T N ) ≤ t , E N W } ≤ C N K t ω + C e − c N .
For arbitrary coefficients, choose a composition series of the natural C m \mathbb{C}^{m} C m -module for their unital algebra and a fixed orthogonal splitting. The coefficients are block upper triangular with r c o m p ≤ m r_{\mathrm{comp}} \le m r comp ≤ m simple diagonal factors. Factors on which all coefficients vanish give identity pencils; the others satisfy the preceding bound on E N W \mathcal{E}_{N}^{\mathrm{W}} E N W .
On E N W \mathcal{E}_{N}^{\mathrm{W}} E N W the strictly upper triangular part has norm O ( N ) O(\sqrt{N}) O ( N ) . Thus the inverse expansion (30) in § 5 gives, when all diagonal least singular values are at least τ f a c ≤ 1 \tau_{\mathrm{fac}} \le1 τ fac ≤ 1 ,
∥ T N − 1 ∥ ≤ C N ( r c o m p − 1 ) / 2 τ f a c − r c o m p , τ f a c = ( 2 C N ( r c o m p − 1 ) / 2 t ) 1 / r c o m p . \left\lVert T_N^{-1}\right\rVert\le CN^{(r_{\mathrm{comp}}-1)/2}\tau_{\mathrm{fac}}^{-r_{\mathrm{comp}}}, \qquad\tau_{\mathrm{fac}} = \left(2CN^{(r_{\mathrm{comp}}-1)/2}t\right)^{1/r_{\mathrm{comp}}}. T N − 1 ≤ C N ( r comp − 1 ) /2 τ fac − r comp , τ fac = ( 2 C N ( r comp − 1 ) /2 t ) 1/ r comp .
For this radius the bound is ( 2 t ) − 1 (2t)^{-1} ( 2 t ) − 1 , so a finite union over the diagonal factors gives C N K ′ t ω / r c o m p + C e − c N CN^{K'}t^{\omega/r_{\mathrm{comp}}} + Ce^{-cN} C N K ′ t ω / r comp + C e − c N on E N W \mathcal{E}_{N}^{\mathrm{W}} E N W . For larger radii the trivial bound suffices; if all coefficients vanish, T N = I T_N = I T N = I . Adding q N W q_N^{\mathrm{W}} q N W once proves the tail.
Finally,
N − 1 ∥ H ℓ ( N ) ∥ H S 2 = N − 2 ( 2 ∑ i < j ∣ ζ ℓ , i j ∣ 2 + ∑ i d i 2 ) . N^{-1}\left\lVert H_{\ell}^{(N)}\right\rVert_{\mathrm{HS}}^{2} = N^{-2}\left(2\sum_{i<j}\lvert\zeta_{\ell,ij}\rvert^{2} + \sum_i d_{i}^{2}\right). N − 1 H ℓ ( N ) HS 2 = N − 2 ( 2 i < j ∑ ∣ ζ ℓ , ij ∣ 2 + i ∑ d i 2 ) .
As in the proof of Theorem 1.2 , the weak law applies to the unordered off-diagonal sum; the diagonal term is negligible by Markov, since its expectation is O ( N − 1 ) O(N^{-1}) O ( N − 1 ) . Thus q N W → 0 q_N^{\mathrm{W}} \to0 q N W → 0 . With ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) moments, apply the estimate for centered sums from that proof separately to these two sums at threshold c N 2 cN^{2} c N 2 . Their normalized mean is 1 + O ( N − 1 ) 1+O(N^{-1}) 1 + O ( N − 1 ) , below the localization threshold, so q N W ≤ C N − min ( ε , 2 ) q_N^{\mathrm{W}} \le CN^{-\min(\varepsilon,2)} q N W ≤ C N − m i n ( ε , 2 ) .
For Gaussian laws, the whitening and chain-rule argument of Proposition 4.4 applies block by block, including the real diagonal atoms. A whitened edge coordinate is still supported on its two endpoints, so its variation has rank at most 2 m 2m 2 m ; diagonal variations have rank at most m m m . The resulting independent real coordinates give ω = 1 \omega= 1 ω = 1 . The exponential Markov argument in the proof of Theorem 1.2 gives q N W ≤ C e − c N 2 q_N^{\mathrm{W}} \le Ce^{-cN^2} q N W ≤ C e − c N 2 . The inverse expansion gives radius exponent 1 / r c o m p 1/r_{\mathrm{comp}} 1/ r comp and total additive error C e − c N Ce^{-cN} C e − c N .
For polynomials in the Hermitian inputs, the word-chain linearization gives the following consequence.
Corollary 6.3 (Polynomial shifts in the Wigner model). For a fixed scalar-coefficient polynomial P P P in the Hermitian variables and every fixed z ≠ P ( 0 ) z \ne P(0) z = P ( 0 ) , finite second moments and the stated bounded densities imply
P { s min ( P ( H ( N ) ) − z I ) ≤ t } ≤ C z N a z t b z + C z e − c z N + q N W \mathbb{P}\{s_{\min}(P(H^{(N)})-zI)\le t\}\le C_zN^{a_z}t^{b_z}+C_ze^{-c_zN}+q_N^{\mathrm{W}} P { s m i n ( P ( H ( N ) ) − z I ) ≤ t } ≤ C z N a z t b z + C z e − c z N + q N W for some fixed positive constants. In particular a polynomial least-singular-value cutoff holds with probability tending to one.
Proof. The constant case is immediate. Otherwise Lemma 2.1, with Z ℓ = H ℓ ( N ) = Z ℓ ∗ Z_\ell=H_\ell^{(N)}=Z_\ell^* Z ℓ = H ℓ ( N ) = Z ℓ ∗ , gives the word-chain pencil
L z ( H ) = K z ⊗ I + ∑ ℓ ( U ℓ + V ℓ ) ⊗ H ℓ ( N ) , det K z = P ( 0 ) − z ≠ 0. L_z(H)=K_z\otimes I+\sum_\ell(U_\ell+V_\ell)\otimes H_\ell^{(N)},\qquad\det K_z=P(0)-z\ne0. L z ( H ) = K z ⊗ I + ℓ ∑ ( U ℓ + V ℓ ) ⊗ H ℓ ( N ) , det K z = P ( 0 ) − z = 0.
Multiplication by K z − 1 ⊗ I K_z^{-1}\otimes I K z − 1 ⊗ I gives a Wigner pencil with identity constant term. The word-chain elimination identity expresses ( P ( H ) − z I ) − 1 (P(H)-zI)^{-1} ( P ( H ) − z I ) − 1 as the upper-left block of L z ( H ) − 1 L_z(H)^{-1} L z ( H ) − 1 , and hence
∥ ( P ( H ) − z I ) − 1 ∥ ≤ ∥ L z ( H ) − 1 ∥ ≤ ∥ [ ( K z − 1 ⊗ I ) L z ( H ) ] − 1 ∥ ∥ K z − 1 ∥ . \lVert(P(H)-zI)^{-1}\rVert\le\lVert L_z(H)^{-1}\rVert\le\lVert[(K_z^{-1}\otimes I)L_z(H)]^{-1}\rVert\lVert K_z^{-1}\rVert. ∥( P ( H ) − z I ) − 1 ∥ ≤ ∥ L z ( H ) − 1 ∥ ≤ ∥[( K z − 1 ⊗ I ) L z ( H ) ] − 1 ∥ ∥ K z − 1 ∥ .
The tail estimate for the normalized pencil proves the assertion. The comparison uses only the fixed matrix K z − 1 K_z^{-1} K z − 1 and requires no norm estimate for the elimination factors. □
Logarithmic potentials and Brown measure convergence
We prove Theorem 1.1 by comparing normalized logarithmic determinants, then take distributional Laplacians to identify the eigenvalue measure. Fix a nonconstant polynomial P P P . For a fixed shift z z z , let ν N z \nu_N^z ν N z be the squared singular-value law of P ( X , X ∗ ) − z I P(X,X^*)-zI P ( X , X ∗ ) − z I , and let ν z \nu^z ν z be the law of ( P ( c , c ∗ ) − z ) ( P ( c , c ∗ ) − z ) ∗ (P(c,c^*)-z)(P(c,c^*)-z)^* ( P ( c , c ∗ ) − z ) ( P ( c , c ∗ ) − z ) ∗ . In the iid argument, P ( Z ) P(Z) P ( Z ) abbreviates P ( Z , Z ∗ ) P(Z,Z^*) P ( Z , Z ∗ ) ; arguments are omitted when the input tuple is fixed. Weak convergence alone does not justify the logarithmic integral. We need both logarithmic tails and convergence after planar integration to identify the Brown measure. We prove:
(i) ν N z \nu_N^z ν N z converges weakly in probability to ν z \nu^z ν z under a finite second moment.
(ii) For z ≠ P ( 0 ) z\ne P(0) z = P ( 0 ) , the logarithmic integrals converge in probability. The finite ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) moment gives a count down to a polynomial scale; Corollary 1.3 controls the interval next to zero. Together they bound the negative logarithmic tail.
(iii) The logarithmic potentials converge in expected local spatial L 1 L^1 L 1 . Taking their distributional Laplacians then gives Brown measure convergence.
We extend the Gaussian and non-Gaussian Hermitization arguments of [15 Sections 5 and 6.2] , based on the logarithmic-potential method of [22 , 7 ] . Rank truncation and Taylor remainder estimates allow the stated weaker moment assumptions for real and correlated planar atoms.
The least singular value estimates in § 3 –§ 5 control the interval closest to zero. The count above that cutoff starts from two results for the free and Gaussian laws. A nonconstant self-adjoint scalar polynomial in free semicircular variables has a non-atomic law and an algebraic Cauchy transform with local Puiseux expansions [21 Theorem 1.1, parts (2), (5), (6), and Corollary 1.2] . For a fixed self-adjoint polynomial in independent normalized GUE matrices, the expected Cauchy transform at i η i\eta i η differs from its free limit by C N − 2 η − b 1 CN^{-2}\eta^{-b_1} C N − 2 η − b 1 for N − b 0 ≤ η ≤ 1 N^{-b_0}\leq\eta\leq1 N − b 0 ≤ η ≤ 1 , for some b 0 , b 1 > 0 b_0,b_1>0 b 0 , b 1 > 0 [3 Lemma 5.5.4] . We apply the latter result only to GUE matrices, writing a standard complex Ginibre matrix as ( H 1 + i H 2 ) / 2 (H_1+iH_2)/\sqrt{2} ( H 1 + i H 2 ) / 2 . We compare the present entry laws with these Gaussian laws below.
Rank truncation and weak singular-value convergence
A rank bound will control the positive logarithmic tail under second moments. Applied to the elimination matrices, it will also transfer the small-value count to the polynomial. The same truncation allows approximation by bounded atoms. For a matrix M M M , let Π T M \Pi_T M Π T M replace each singular value s s s by min ( s , T ) \min(s,T) min ( s , T ) . This is the Hilbert–Schmidt projection onto the convex operator-norm ball. Hence
∥ Π T M − Π T M ′ ∥ H S ≤ ∥ M − M ′ ∥ H S , rank ( M − Π T M ) ≤ T − 2 ∥ M ∥ H S 2 . \lVert\Pi_T M-\Pi_T M'\rVert_{\mathrm{HS}}\leq\lVert M-M'\rVert_{\mathrm{HS}},\qquad\operatorname{rank}(M-\Pi_T M)\leq T^{-2}\lVert M\rVert_{\mathrm{HS}}^2. ∥ Π T M − Π T M ′ ∥ HS ≤ ∥ M − M ′ ∥ HS , rank ( M − Π T M ) ≤ T − 2 ∥ M ∥ HS 2 .
For Hermitian M M M , this clips the eigenvalues to [ − T , T ] [-T,T] [ − T , T ] . We apply the truncation separately to each input matrix.
Lemma 7.1 (Upper singular-value count from second moments). Suppose Z = ( Z 1 , … , Z n ) Z=(Z_1,\ldots,Z_n) Z = ( Z 1 , … , Z n ) satisfies E ∥ Z ℓ ∥ H S 2 ≤ C N \mathbb{E}\lVert Z_\ell\rVert_{\mathrm{HS}}^2\leq CN E ∥ Z ℓ ∥ HS 2 ≤ C N . If Q Q Q is a fixed square matrix-valued ∗ \ast ∗ -polynomial of degree at most D ≥ 1 D\geq1 D ≥ 1 , whose coefficient matrices have fixed dimensions, then
E # { j : s j ( Q ( Z ) ) > M } ≤ C Q N M − 2 / D , M ≥ 1. (44) \mathbb{E}\#\{j:s_j(Q(Z))>M\}\leq C_QNM^{-2/D},\qquad M\geq1. \tag*{(44)} E # { j : s j ( Q ( Z )) > M } ≤ C Q N M − 2/ D , M ≥ 1. (44) The constants may depend on the fixed coefficient dimensions and the second-moment bound.
Proof. For each word, replacing its factors one at a time gives
rank ( Q ( Z ) − Q ( Z ′ ) ) ≤ C Q ∑ ℓ rank ( Z ℓ − Z ℓ ′ ) . \operatorname{rank}(Q(Z)-Q(Z'))\leq C_Q\sum_\ell\operatorname{rank}(Z_\ell-Z'_\ell). rank ( Q ( Z ) − Q ( Z ′ )) ≤ C Q ℓ ∑ rank ( Z ℓ − Z ℓ ′ ) .
Indeed, telescoping a word gives summands with a factor Z ℓ − Z ℓ ′ Z_\ell-Z'_\ell Z ℓ − Z ℓ ′ or its adjoint, and multiplication cannot increase rank. With Z ′ = Π T Z Z'=\Pi_TZ Z ′ = Π T Z , it follows that
E rank ( Q ( Z ) − Q ( Π T Z ) ) ≤ C Q N T − 2 , ∥ Q ( Π T Z ) ∥ ≤ C Q max ( 1 , T ) D . \mathbb{E}\operatorname{rank}(Q(Z)-Q(\Pi_TZ))\leq C_QNT^{-2},\qquad\lVert Q(\Pi_TZ)\rVert\leq C_Q\max(1,T)^D. E rank ( Q ( Z ) − Q ( Π T Z )) ≤ C Q N T − 2 , ∥ Q ( Π T Z )∥ ≤ C Q max ( 1 , T ) D .
For all sufficiently large M M M , choose T = c Q M 1 / D T=c_QM^{1/D} T = c Q M 1/ D so that the last norm is at most M M M . Rank interlacing then bounds the number of singular values of Q ( Z ) Q(Z) Q ( Z ) above M M M by the rank difference. This proves (44) for large M M M ; increasing the constant gives the remaining values M ≥ 1 M\geq1 M ≥ 1 . The estimate also holds for constant polynomials. □
For assertion (i), suppose first that the atoms are bounded, centered, and have absolute variance one. Expand the normalized trace of a fixed ∗ \ast ∗ -word of length k k k as a sum over closed index walks. Each term has normalization N − 1 − k / 2 N^{-1-k/2} N − 1 − k /2 . If it involves e e e distinct atoms ( ℓ , a , b ) (\ell,a,b) ( ℓ , a , b ) and v v v spatial indices, centering and connectedness imply
e ≤ k / 2 , v ≤ e + 1 ≤ 1 + k / 2. e\leq k/2,\qquad v\leq e+1\leq1+k/2. e ≤ k /2 , v ≤ e + 1 ≤ 1 + k /2.
There are finitely many equality patterns for fixed k k k , and boundedness controls the atom moments in their contributions. Therefore only patterns with v = 1 + k / 2 v=1+k/2 v = 1 + k /2 can have a nonzero limit.
Equality forces e = k / 2 e=k/2 e = k /2 and the spatial graph to be a tree. Loops, parallel edges with different matrix labels, or counting the reverse entry as a separate atom would reduce the number of vertices. Each edge of the tree is traversed exactly twice, once in each direction. For the two factors to use the same directed atom, they must have opposite star status, so their expectation is E ( ξ ξ ˉ ) = 1 \mathbb{E}(\xi\bar{\xi})=1 E ( ξ ξ ˉ ) = 1 . The number of distinct vertex labelings is N ( N − 1 ) ⋯ ( N − v + 1 ) N(N-1)\cdots(N-v+1) N ( N − 1 ) ⋯ ( N − v + 1 ) , and its normalized contribution tends to one. Leaf removal gives precisely the circular noncrossing pairings with matching matrix indices and opposite star status. A pairing involving E ξ 2 = ρ E_{\xi}^{2}=\rho E ξ 2 = ρ or its conjugate would have equal star status and hence cannot occur in a leading tree pattern. Every remaining nonzero pattern is O ( N − 1 / 2 ) O(N^{-1/2}) O ( N − 1/2 ) .
In the covariance of two normalized traces, patterns with disjoint atom sets cancel. A shared atom joins the two underlying graphs, and a surviving pattern of total length k k k still has v ≤ e + 1 ≤ 1 + k / 2 v \le e + 1 \le1 + k/2 v ≤ e + 1 ≤ 1 + k /2 . Its normalization is now N − 2 − k / 2 N^{-2-k/2} N − 2 − k /2 , so its contribution is O ( N − 1 ) O(N^{-1}) O ( N − 1 ) . The variance therefore tends to zero, and the joint ∗ * ∗ -moments converge in probability. Applying this convergence to powers of ( P − z ) ( P − z ) ∗ (P-z)(P-z)^* ( P − z ) ( P − z ) ∗ gives convergence of every moment of ν z N \nu_{z}^{N} ν z N in probability. To pass to weak convergence, take an arbitrary subsequence and a further subsequence along which all these moments converge almost surely. The next higher moments give tightness and uniform integrability of each lower moment. Every subsequential weak limit therefore has the moments of ν z \nu^{z} ν z . Since ν z \nu^{z} ν z has compact support, moment determinacy gives weak convergence in probability.
For atoms with finite second moments, define the bounded approximants by
a L = E { ξ 1 { ∣ ξ ∣ ≤ L } } , b L 2 = E { ∣ ξ 1 { ∣ ξ ∣ ≤ L } − a L ∣ 2 } , ξ L = b L − 1 ( ξ 1 { ∣ ξ ∣ ≤ L } − a L ) . a_L=\mathbb{E}\{\xi\mathbf{1}_{\{|\xi|\le L\}}\},\qquad b_L^2=\mathbb{E}\{\lvert\xi\mathbf{1}_{\{|\xi|\le L\}}-a_L\rvert^2\},\qquad\xi_L=b_L^{-1}\bigl(\xi\mathbf{1}_{\{|\xi|\le L\}}-a_L\bigr). a L = E { ξ 1 { ∣ ξ ∣ ≤ L } } , b L 2 = E {∣ ξ 1 { ∣ ξ ∣ ≤ L } − a L ∣ 2 } , ξ L = b L − 1 ( ξ 1 { ∣ ξ ∣ ≤ L } − a L ) .
For all sufficiently large L L L , b L > 0 b_L>0 b L > 0 , and
a L ⟶ 0 , b L ⟶ 1 , δ L 2 : = E ∣ ξ − ξ L ∣ 2 ⟶ 0. a_L\longrightarrow0,\qquad b_L\longrightarrow1,\qquad\delta_L^2:=\mathbb{E}|\xi-\xi_L|^2\longrightarrow0. a L ⟶ 0 , b L ⟶ 1 , δ L 2 := E ∣ ξ − ξ L ∣ 2 ⟶ 0.
The bounded law ξ L \xi_L ξ L is centered and has absolute variance one. Its pseudovariance may differ from that of ξ \xi ξ ; the preceding moment argument applies in either case.
Couple X X X and Y L Y_L Y L entrywise through ξ \xi ξ and ξ L \xi_L ξ L . This changes most entries, so their difference need not have small rank. Insert operator-norm truncations between the two polynomials and compare
P ( X ) − z I , P ( Π T X ) − z I , P ( Π T Y L ) − z I , P ( Y L ) − z I . P(X)-zI,\qquad P(\Pi_T X)-zI,\qquad P(\Pi_TY_L)-zI,\qquad P(Y_L)-zI. P ( X ) − z I , P ( Π T X ) − z I , P ( Π T Y L ) − z I , P ( Y L ) − z I .
The two outer differences have expected rank at most C P N T − 2 C_PNT^{-2} C P N T − 2 . For the middle difference, telescope each word and use the norm bounds on the truncated inputs and nonexpansiveness of Π T \Pi_T Π T to obtain
E N − 1 ∥ P ( Π T X ) − P ( Π T Y L ) ∥ H S 2 ≤ C P , z ( T ) ∑ ℓ E N − 1 ∥ X ℓ − Y ℓ , L ∥ H S 2 ≤ C P , z ( T ) δ L 2 . \mathbb{E}N^{-1}\lVert P(\Pi_TX)-P(\Pi_TY_L)\rVert_{\mathrm{HS}}^2
\le C_{P,z}(T)\sum_{\ell}\mathbb{E}N^{-1}\lVert X_\ell-Y_{\ell,L}\rVert_{\mathrm{HS}}^2
\le C_{P,z}(T)\delta_L^2. E N − 1 ∥ P ( Π T X ) − P ( Π T Y L ) ∥ HS 2 ≤ C P , z ( T ) ℓ ∑ E N − 1 ∥ X ℓ − Y ℓ , L ∥ HS 2 ≤ C P , z ( T ) δ L 2 .
Both middle outputs have norm at most a fixed C P , z ( T ) C_{P,z}(T) C P , z ( T ) . Combining Mirsky’s inequality and Cauchy–Schwarz with ∣ s 2 − t 2 ∣ ≤ 2 C P , z ( T ) ∣ s − t ∣ |s^2-t^2|\le2C_{P,z}(T)|s-t| ∣ s 2 − t 2 ∣ ≤ 2 C P , z ( T ) ∣ s − t ∣ bounds the difference of their squared singular-value laws. Write ν Z , N z \nu_{Z,N}^{z} ν Z , N z for the squared singular-value law with input Z Z Z . Then, for f ∈ C c 1 ( R ) f\in C_c^1(\mathbb{R}) f ∈ C c 1 ( R ) ,
E ∣ ∫ f d ν X , N z − ∫ f d ν Y L , N z ∣ ≤ C f T − 2 + C f , P , z ( T ) δ L . (45) \mathbb{E}\left|\int f\,d\nu_{X,N}^{z}-\int f\,d\nu_{Y_L,N}^{z}\right|
\le C_fT^{-2}+C_{f,P,z}(T)\delta_L. \tag*{(45)} E ∫ f d ν X , N z − ∫ f d ν Y L , N z ≤ C f T − 2 + C f , P , z ( T ) δ L . (45)
For the outer comparisons, integrate f ′ f' f ′ against the distribution functions and use rank interlacing. For the middle comparison, use
1 N ∑ j ∣ s j 2 − t j 2 ∣ ≤ 2 C P , z ( T ) ( 1 N ∑ j ∣ s j − t j ∣ 2 ) 1 / 2 ≤ 2 C P , z ( T ) N − 1 / 2 ∥ A − B ∥ H S , \frac{1}{N}\sum_j|s_j^2-t_j^2|
\le2C_{P,z}(T)\left(\frac{1}{N}\sum_j|s_j-t_j|^2\right)^{1/2}
\le2C_{P,z}(T)N^{-1/2}\lVert A-B\rVert_{\mathrm{HS}}, N 1 j ∑ ∣ s j 2 − t j 2 ∣ ≤ 2 C P , z ( T ) ( N 1 j ∑ ∣ s j − t j ∣ 2 ) 1/2 ≤ 2 C P , z ( T ) N − 1/2 ∥ A − B ∥ HS ,
where A , B A,B A , B are the middle outputs. This proves (45) . Choose first T T T large and then L L L large, and apply the bounded-law limit with this fixed L L L as N → ∞ N\to\infty N → ∞ . The upper count (44) gives tightness, so these compactly supported test functions suffice, proving assertion (i). To integrate the negative logarithm, we next need a quantitative bound on mass near zero.
Small singular values for the free and Gaussian laws
For assertion (ii), assume now that E ∣ ξ ∣ 2 + ε < ∞ \mathbb{E}|\xi|^{2+\varepsilon}<\infty E ∣ ξ ∣ 2 + ε < ∞ for some ε > 0 \varepsilon>0 ε > 0 . We need a power bound for the expected small-singular-value count down to a polynomial scale. We first prove it for the free law and the Gaussian polynomial, then for its linearization. The pencil and elimination matrices are evaluated as matrix-valued ∗ * ∗ -polynomials: L z ( G ) \mathcal{L}_z(G) L z ( G ) uses G ∗ G^* G ∗ , and on real inputs it agrees with the transpose pencil.
Write each circular variable as c ℓ = ( s ℓ , 1 + i s ℓ , 2 ) / 2 c_\ell=(s_{\ell,1}+is_{\ell,2})/\sqrt{2} c ℓ = ( s ℓ , 1 + i s ℓ , 2 ) / 2 . The scalar self-adjoint polynomial Q z = ( P − z ) ( P − z ) ∗ Q_z=(P-z)(P-z)^* Q z = ( P − z ) ( P − z ) ∗ in these semicircular generators is nonconstant. Indeed, its highest-degree term is P d P d ∗ ≠ 0 P_dP_d^*\ne0 P d P d ∗ = 0 in the free algebra. The change from x ℓ , x ℓ ∗ x_\ell,x_\ell^* x ℓ , x ℓ ∗ to s ℓ , 1 , s ℓ , 2 s_{\ell,1},s_{\ell,2} s ℓ , 1 , s ℓ , 2 is invertible, and semicircular variables satisfy no polynomial relations. Thus the nonconstant scalar case of [21 ] applies. Use the convention
g z ( w ) = ∫ ( x − w ) − 1 ν z ( d x ) , g^z(w)=\int(x-w)^{-1}\nu^z(dx), g z ( w ) = ∫ ( x − w ) − 1 ν z ( d x ) ,
whose imaginary part is positive in the upper half-plane. The non-atomicity and the local Puiseux expansion imply that, for some θ > 0 \theta>0 θ > 0 ,
ν z ( [ 0 , x ] ) ≤ C x θ , 0 < x ≤ x ∗ . (46) \nu^z([0,x])\le Cx^\theta,\qquad0<x\le x_*. \tag*{(46)} ν z ([ 0 , x ]) ≤ C x θ , 0 < x ≤ x ∗ . (46)
Indeed, the Cauchy bound ∣ g z ( i η ) ∣ ≤ η − 1 |g^z(i\eta)|\le\eta^{-1} ∣ g z ( i η ) ∣ ≤ η − 1 excludes Puiseux powers below − 1 -1 − 1 , and the absence of an atom at zero excludes a nonzero w − 1 w^{-1} w − 1 term. After decreasing θ \theta θ to be less than 1 1 1 , it follows that Im g z ( i η ) ≤ C η θ − 1 \operatorname{Im}g^z(i\eta)\le C\eta^{\theta-1} Im g z ( i η ) ≤ C η θ − 1 . Since
ν z ( [ 0 , x ] ) ≤ 2 x Im g z ( i x ) , \nu^z([0,x])\le2x\operatorname{Im}g^z(ix), ν z ([ 0 , x ]) ≤ 2 x Im g z ( i x ) ,
we obtain (46) . In particular the negative logarithm is integrable against ν z \nu^z ν z .
Let G = ( G 1 , … , G n ) G=(G_1,\ldots,G_n) G = ( G 1 , … , G n ) be a family of independent normalized complex Ginibre matrices. Apply the GUE resolvent estimate to Q z Q_z Q z , using their representation by pairs of independent GUE matrices. Choose 0 < a 0 < b 0 0<a_0<b_0 0 < a 0 < b 0 sufficiently small that N − 2 η − b 1 ≤ C η θ − 1 N^{-2}\eta^{-b_1}\le C\eta^{\theta-1} N − 2 η − b 1 ≤ C η θ − 1 for η ≥ N − a 0 \eta\ge N^{-a_0} η ≥ N − a 0 . The preceding mass inequality gives
E N − 1 # { j : s j ( P ( G , G ∗ ) − z I ) 2 ≤ x } ≤ C x θ , N − a 0 ≤ x ≤ 1. (47) \mathbb{E}N^{-1}\#\{j:s_j(P(G,G^*)-zI)^2\le x\}\le Cx^\theta,\qquad N^{-a_0}\le x\le1. \tag*{(47)} E N − 1 # { j : s j ( P ( G , G ∗ ) − z I ) 2 ≤ x } ≤ C x θ , N − a 0 ≤ x ≤ 1. (47)
We have enlarged C C C to include x x x away from zero.
To compare entry laws, we use the affine pencil: one atom changes it by bounded rank, which controls resolvent derivatives. We first transfer the Gaussian count to this pencil. Let h h h be its coefficient size and define its Hermitization and expected Cauchy transform by
H z ( X ) = ( 0 L z ( X ) L z ( X ) ∗ 0 ) , m X , N ( w ) = E tr 2 h N ( H z ( X ) − w ) − 1 . \mathcal{H}_z(X)=
\begin{pmatrix}
0 & \mathcal{L}_z(X)\\
\mathcal{L}_z(X)^* & 0
\end{pmatrix},
\qquad
m_{X,N}(w)=\mathbb{E}\operatorname{tr}_{2hN}\bigl(\mathcal{H}_z(X)-w\bigr)^{-1}. H z ( X ) = ( 0 L z ( X ) ∗ L z ( X ) 0 ) , m X , N ( w ) = E tr 2 h N ( H z ( X ) − w ) − 1 .
The Gaussian norm estimate follows from a net argument. A 1 / 4 1/4 1/4 -net of the unit sphere of C N \mathbb{C}^N C N has at most 9 2 N 9^{2N} 9 2 N points, and nets for both vectors give
∥ G ℓ ∥ ≤ 2 max u , v in the nets ∣ u ∗ G ℓ v ∣ . \|G_\ell\|\le2\max_{u,v\ \text{in the nets}}|u^*G_\ell v|. ∥ G ℓ ∥ ≤ 2 u , v in the nets max ∣ u ∗ G ℓ v ∣.
For fixed unit vectors, u ∗ G ℓ v u^*G_\ell v u ∗ G ℓ v is proper complex Gaussian of variance 1 / N 1/N 1/ N , with tail e − N t 2 e^{-Nt^2} e − N t 2 . A union bound over at most 9 4 N 9^{4N} 9 4 N pairs and the fixed number of input matrices shows that
P { max ℓ ∥ G ℓ ∥ > C 0 } ≤ C e − c N \mathbb{P}\{\max_\ell\|G_\ell\|>C_0\}\le Ce^{-cN} P { ℓ max ∥ G ℓ ∥ > C 0 } ≤ C e − c N
for sufficiently large fixed C 0 C_0 C 0 . On the complementary event, V z , W z \mathcal{V}_z,\mathcal{W}_z V z , W z and their inverses have fixed norm bounds. Set Δ z = diag ( P ( G , G ∗ ) − z I , I ) \Delta_z=\operatorname{diag}(P(G,G^*)-zI,I) Δ z = diag ( P ( G , G ∗ ) − z I , I ) . The identity Δ z = V z Z z W z \Delta_z=\mathcal{V}_z\mathcal{Z}_z\mathcal{W}_z Δ z = V z Z z W z and min–max give s j ( Δ z ) ≤ C 1 s j ( L z ) s_j(\Delta_z) \le C_1s_j(\mathcal{L}_z) s j ( Δ z ) ≤ C 1 s j ( L z ) . For C 1 t < 1 C_1t < 1 C 1 t < 1 , the identity block of Δ z \Delta_z Δ z contributes no singular values at most C 1 t C_1t C 1 t , and hence
# { s j ( L z ) ≤ t } ≤ # { s j ( P ( G , G ∗ ) − z I ) ≤ C 1 t } . \#\{s_j(\mathcal{L}_z) \le t\} \le\#\{s_j(P(G,G^*)-zI) \le C_1t\}. # { s j ( L z ) ≤ t } ≤ # { s j ( P ( G , G ∗ ) − z I ) ≤ C 1 t } .
Hermitization counts every singular value twice and normalizes by 2 h N 2hN 2 h N . Thus (47) bounds its expected mass in [ − t , t ] [-t,t] [ − t , t ] by C t 2 θ + C e − c N Ct^{2\theta}+Ce^{-cN} C t 2 θ + C e − c N whenever ( C 1 t ) 2 ≥ N − a 0 (C_1t)^2 \ge N^{-a_0} ( C 1 t ) 2 ≥ N − a 0 and t t t is small. We may therefore choose 0 < κ < min ( 2 θ , 1 ) 0 < \kappa< \min(2\theta,1) 0 < κ < min ( 2 θ , 1 ) and 0 < a 1 < a 0 / 2 0 < a_1 < a_0/2 0 < a 1 < a 0 /2 so that this expected mass is at most C t κ Ct^\kappa C t κ for N − a 1 ≤ t ≤ 1 N^{-a_1} \le t \le1 N − a 1 ≤ t ≤ 1 .
For η ≥ N − a 1 \eta\ge N^{-a_1} η ≥ N − a 1 , integrate the kernel η / ( λ 2 + η 2 ) \eta/(\lambda^2+\eta^2) η / ( λ 2 + η 2 ) first on ∣ λ ∣ ≤ η |\lambda| \le\eta ∣ λ ∣ ≤ η , and then on the shells 2 j η < ∣ λ ∣ ≤ 2 j + 1 η ≤ 1 2^j\eta< |\lambda| \le2^{j+1}\eta\le1 2 j η < ∣ λ ∣ ≤ 2 j + 1 η ≤ 1 . The first interval contributes at most C η κ − 1 C\eta^{\kappa-1} C η κ − 1 . On the j j j th shell the contribution is at most
η ( 2 j η ) 2 C ( 2 j + 1 η ) κ ≤ C η κ − 1 2 j ( κ − 2 ) . \frac{\eta}{(2^j\eta)^2}C(2^{j+1}\eta)^\kappa\le C\eta^{\kappa-1}2^{j(\kappa-2)}. ( 2 j η ) 2 η C ( 2 j + 1 η ) κ ≤ C η κ − 1 2 j ( κ − 2 ) .
The region bounded away from zero contributes O ( η ) O(\eta) O ( η ) ; a possible last shell meeting 1 changes only the constant. Summing gives
ℑ m G , N ( i η ) ≤ C η κ − 1 , N − a 1 ≤ η ≤ 1. (48) \Im m_{G,N}(i\eta) \le C\eta^{\kappa-1}, \qquad N^{-a_1} \le\eta\le1. \tag*{(48)} ℑ m G , N ( i η ) ≤ C η κ − 1 , N − a 1 ≤ η ≤ 1. (48)
The counting bounds apply throughout the summation, since all scales are at least η \eta η .
Resolvent comparison for general entry laws
We transfer the Gaussian pencil bound in two steps. Replacing each atom by a Gaussian with the same full real covariance uses the fractional moment assumption; Gaussian interpolation then removes the pseudovariance. Both errors must be small enough at the counting scale. Write the Hermitization as
H z ( X ) = D z ⊗ I + 1 N ∑ ℓ , a , b ( ξ ℓ , a b C ℓ ⊗ E a b + ξ ‾ ℓ , a b C ℓ ∗ ⊗ E b a ) , C ℓ = ( 0 U ℓ V ℓ ∗ 0 ) . (49) \mathcal{H}_z(X) = D_z \otimes I + \frac{1}{\sqrt{N}}\sum_{\ell,a,b}\left(\xi_{\ell,ab}C_\ell\otimes E_{ab}+\overline{\xi}_{\ell,ab}C_\ell^* \otimes E_{ba}\right), \qquad C_\ell=\begin{pmatrix}0&U_\ell\\V_\ell^*&0\end{pmatrix}. \tag*{(49)} H z ( X ) = D z ⊗ I + N 1 ℓ , a , b ∑ ( ξ ℓ , ab C ℓ ⊗ E ab + ξ ℓ , ab C ℓ ∗ ⊗ E ba ) , C ℓ = ( 0 V ℓ ∗ U ℓ 0 ) . (49)
where D z = ( 0 K z K z ∗ 0 ) D_z=\begin{pmatrix}0&K_z\\K_z^*&0\end{pmatrix} D z = ( 0 K z ∗ K z 0 ) . Indeed, the upper-right blocks contributed by one atom are ξ U ℓ E a b \xi U_\ell E_{ab} ξ U ℓ E ab and ξ ‾ V ℓ E b a \overline{\xi}V_\ell E_{ba} ξ V ℓ E ba , the two terms of L z \mathcal{L}_z L z ; the lower-left blocks are their adjoints.
Let γ ρ \gamma_\rho γ ρ be a centered complex Gaussian with real covariance (9) , and let G ρ G_\rho G ρ have independent entries γ ρ / N \gamma_\rho/\sqrt{N} γ ρ / N . Thus E ∣ γ ρ ∣ 2 = 1 \mathbb{E}|\gamma_\rho|^2=1 E ∣ γ ρ ∣ 2 = 1 and E γ ρ 2 = ρ \mathbb{E}\gamma_\rho^2=\rho E γ ρ 2 = ρ . The choice ρ = 1 \rho=1 ρ = 1 is a standard real Gaussian, and G = G 0 G=G_0 G = G 0 is the standard complex Ginibre family. We compare X X X with G ρ G_\rho G ρ , then G ρ G_\rho G ρ with G G G .
Lemma 7.2 (Fractional comparison of whole atoms). For 0 < α ≤ 1 0 < \alpha\le1 0 < α ≤ 1 , if E ∣ ξ ∣ 2 + α < ∞ \mathbb{E}|\xi|^{2+\alpha}<\infty E ∣ ξ ∣ 2 + α < ∞ and ℑ w > 0 \Im w > 0 ℑ w > 0 , then, in either input class,
∣ m X , N ( w ) − m G ρ , N ( w ) ∣ ≤ C N − α / 2 ( ℑ w ) − 3 − α . |m_{X,N}(w)-m_{G_\rho,N}(w)| \le CN^{-\alpha/2}(\Im w)^{-3-\alpha}. ∣ m X , N ( w ) − m G ρ , N ( w ) ∣ ≤ C N − α /2 ( ℑ w ) − 3 − α . No density assumption is needed for this lemma.
Proof. Replace the n N 2 nN^2 n N 2 independent entries one at a time, treating the real and imaginary parts of each entry as a single atom. At each replacement, condition on the other entries. The Hermitian background H ( 0 ) H^{(0)} H ( 0 ) is then independent of the selected atom. For v = ( v 1 , v 2 ) v=(v_1,v_2) v = ( v 1 , v 2 ) and χ = v 1 + i v 2 \chi=v_1+iv_2 χ = v 1 + i v 2 , write its perturbation as
A ( v ) = N − 1 / 2 ( χ C ℓ ⊗ E a b + χ ‾ C ℓ ∗ ⊗ E b a ) . A(v)=N^{-1/2}\left(\chi C_\ell\otimes E_{ab}+\overline{\chi}C_\ell^*\otimes E_{ba}\right). A ( v ) = N − 1/2 ( χ C ℓ ⊗ E ab + χ C ℓ ∗ ⊗ E ba ) .
It is Hermitian, has rank at most 4 h 4h 4 h , and satisfies ∥ A ( v ) ∥ ≤ C ∥ v ∥ / N \|A(v)\|\le C\|v\|/\sqrt{N} ∥ A ( v ) ∥ ≤ C ∥ v ∥/ N . For real atoms take v 2 = 0 v_2=0 v 2 = 0 . Set f v ( t ) = tr 2 h N ( H ( 0 ) + t A ( v ) − w ) − 1 f_v(t)=\operatorname{tr}_{2hN}(H^{(0)}+tA(v)-w)^{-1} f v ( t ) = tr 2 h N ( H ( 0 ) + t A ( v ) − w ) − 1 and η = ℑ w \eta=\Im w η = ℑ w . With R t = ( H ( 0 ) + t A ( v ) − w ) − 1 R_t=(H^{(0)}+tA(v)-w)^{-1} R t = ( H ( 0 ) + t A ( v ) − w ) − 1 , differentiation gives
f v ( j ) ( t ) = ( − 1 ) j j ! tr 2 h N [ R t ( A ( v ) R t ) j ] . (50) f_v^{(j)}(t)=(-1)^j j!\operatorname{tr}_{2hN}\left[R_t(A(v)R_t)^j\right]. \tag*{(50)} f v ( j ) ( t ) = ( − 1 ) j j ! tr 2 h N [ R t ( A ( v ) R t ) j ] . (50)
For j ≥ 1 j \ge1 j ≥ 1 , the product has bounded rank and ∥ R t ∥ ≤ η − 1 \lVert R_t\rVert\le\eta^{-1} ∥ R t ∥ ≤ η − 1 . The normalized trace contributes a factor N − 1 N^{-1} N − 1 , so
∣ f v ′ ′ ( t ) ∣ ≤ C N − 2 η − 3 ∥ v ∥ 2 , ∣ f v ′ ′ ′ ( t ) ∣ ≤ C N − 5 / 2 η − 4 ∥ v ∥ 3 . \lvert f''_v(t)\rvert\le CN^{-2}\eta^{-3}\lVert v\rVert^2,\qquad\lvert f'''_v(t)\rvert\le CN^{-5/2}\eta^{-4}\lVert v\rVert^3. ∣ f v ′′ ( t )∣ ≤ C N − 2 η − 3 ∥ v ∥ 2 , ∣ f v ′′′ ( t )∣ ≤ C N − 5/2 η − 4 ∥ v ∥ 3 .
The second-order Taylor remainder in the two real coordinates is
Rem 2 ( v ) = f v ( 1 ) − f v ( 0 ) − f v ′ ( 0 ) − 1 2 f v ′ ′ ( 0 ) = ∫ 0 1 ( 1 − t ) ( f v ′ ′ ( t ) − f v ′ ′ ( 0 ) ) d t . \operatorname{Rem}_2(v)=f_v(1)-f_v(0)-f'_v(0)-\frac{1}{2}f''_v(0)=\int_0^1(1-t)(f''_v(t)-f''_v(0))\,dt. Rem 2 ( v ) = f v ( 1 ) − f v ( 0 ) − f v ′ ( 0 ) − 2 1 f v ′′ ( 0 ) = ∫ 0 1 ( 1 − t ) ( f v ′′ ( t ) − f v ′′ ( 0 )) d t .
Bounding the integrand by the second derivatives, or integrating the third derivative, gives
∣ Rem 2 ( v ) ∣ ≤ C min { N − 2 η − 3 ∥ v ∥ 2 , N − 5 / 2 η − 4 ∥ v ∥ 3 } ≤ C N − 2 − α / 2 η − 3 − α ∥ v ∥ 2 + α . \begin{aligned}
\lvert\operatorname{Rem}_2(v)\rvert\le C\min\left\{N^{-2}\eta^{-3}\lVert v\rVert^2,N^{-5/2}\eta^{-4}\lVert v\rVert^3\right\} \\
&\le CN^{-2-\alpha/2}\eta^{-3-\alpha}\lVert v\rVert^{2+\alpha}.
\end{aligned} ∣ Rem 2 ( v )∣ ≤ C min { N − 2 η − 3 ∥ v ∥ 2 , N − 5/2 η − 4 ∥ v ∥ 3 } ≤ C N − 2 − α /2 η − 3 − α ∥ v ∥ 2 + α .
where the last step uses min ( a , b ) ≤ a 1 − α b α \min(a,b)\le a^{1-\alpha}b^\alpha min ( a , b ) ≤ a 1 − α b α . The constant term does not depend on v v v ; the next two terms are respectively linear in v 1 , v 2 v_1,v_2 v 1 , v 2 and quadratic in these coordinates. Their expectations agree for ξ \xi ξ and γ ρ \gamma_\rho γ ρ , because the means and the entire real covariance agree, including the mixed v 1 v 2 v_1v_2 v 1 v 2 entry. This remains true for the rank-one covariance of class (A). Summing the expected remainders over n N 2 nN^2 n N 2 replacements proves the claim. Every intermediate background is Hermitian, so the resolvent bound holds regardless of the other entries. □
The following covariance comparison develops the argument of [15 Lemma 6.11] .
Lemma 7.3 (Removing arbitrary Gaussian pseudovariance). For ∣ ρ ∣ ≤ 1 \lvert\rho\rvert\le1 ∣ ρ ∣ ≤ 1 and ℑ w > 0 \Im w>0 ℑ w > 0 ,
∣ m G ρ , N ( w ) − m G , N ( w ) ∣ ≤ C N − 1 ( ℑ w ) − 3 . \lvert m_{G_\rho,N}(w)-m_{G,N}(w)\rvert\le CN^{-1}(\Im w)^{-3}. ∣ m G ρ , N ( w ) − m G , N ( w )∣ ≤ C N − 1 ( ℑ w ) − 3 . The constant may be chosen uniformly over ∣ ρ ∣ ≤ 1 \lvert\rho\rvert\le1 ∣ ρ ∣ ≤ 1 .
Proof. Let Y ρ , Y 0 Y_\rho,Y_0 Y ρ , Y 0 denote the centered Gaussian parts of (49) . For deterministic M M M , define the fixed linear covariance maps K ρ ( M ) = E [ Y ρ M Y ρ ] K_\rho(M)=\mathbb{E}[Y_\rho M Y_\rho] K ρ ( M ) = E [ Y ρ M Y ρ ] and K 0 ( M ) = E [ Y 0 M Y 0 ] K_0(M)=\mathbb{E}[Y_0 M Y_0] K 0 ( M ) = E [ Y 0 M Y 0 ] . We apply these maps pointwise when M M M is random. Terms involving distinct entries vanish by independence and centering. The two terms pairing an entry with its conjugate agree because the absolute variances are one. The other two terms give
( K ρ − K 0 ) ( M ) = 1 N ∑ ℓ , a , b [ ρ ( C ℓ ⊗ E a b ) M ( C ℓ ⊗ E a b ) + ρ ‾ ( C ℓ ∗ ⊗ E b a ) M ( C ℓ ∗ ⊗ E b a ) ] . (51) (K_\rho-K_0)(M)=\frac{1}{N}\sum_{\ell,a,b}\left[\rho(C_\ell\otimes E_{ab})M(C_\ell\otimes E_{ab})+\overline{\rho}(C_\ell^*\otimes E_{ba})M(C_\ell^*\otimes E_{ba})\right]. \tag*{(51)} ( K ρ − K 0 ) ( M ) = N 1 ℓ , a , b ∑ [ ρ ( C ℓ ⊗ E ab ) M ( C ℓ ⊗ E ab ) + ρ ( C ℓ ∗ ⊗ E ba ) M ( C ℓ ∗ ⊗ E ba ) ] . (51)
For an N × N N\times N N × N matrix Q Q Q , ∑ a , b E a b Q E a b = Q T \sum_{a,b}E_{ab}QE_{ab}=Q^{\mathsf T} ∑ a , b E ab Q E ab = Q T . Thus the sums in (7.7) transpose each spatial block of M M M , with fixed coefficient matrices on either side. Every such block has norm at most ∥ M ∥ \lVert M\rVert ∥ M ∥ , transpose preserves its norm, and a 2 h × 2 h 2h\times2h 2 h × 2 h block array has norm at most 2 h 2h 2 h times its largest block norm. It follows that
∥ ( K ρ − K 0 ) ( M ) ∥ ≤ C N − 1 ∥ M ∥ , (52) \lVert(K_\rho-K_0)(M)\rVert\le CN^{-1}\lVert M\rVert, \tag*{(52)} ∥( K ρ − K 0 ) ( M )∥ ≤ C N − 1 ∥ M ∥ , (52)
uniformly in ∣ ρ ∣ ≤ 1 \lvert\rho\rvert\le1 ∣ ρ ∣ ≤ 1 .
Interpolate independent copies by H t = D z ⊗ I + t Y ρ + 1 − t Y 0 H_t=D_z\otimes I+\sqrt{t}Y_\rho+\sqrt{1-t}Y_0 H t = D z ⊗ I + t Y ρ + 1 − t Y 0 , and write R t = ( H t − w ) − 1 R_t=(H_t-w)^{-1} R t = ( H t − w ) − 1 . Gaussian covariance differentiation gives
d d t E R t = E [ R t ( K ρ − K 0 ) ( R t ) R t ] . \frac{d}{dt}\mathbb{E}R_t=\mathbb{E}\left[R_t(K_\rho-K_0)(R_t)R_t\right]. d t d E R t = E [ R t ( K ρ − K 0 ) ( R t ) R t ] .
To verify the identity, choose a real Hermitian coefficient basis H r H_r H r and let Δ r s \Delta_{rs} Δ r s be the difference of its real covariances. The derivative of the Gaussian expectation is 1 2 ∑ r , s Δ r s E ∂ r ∂ s R \frac{1}{2}\sum_{r,s}\Delta_{rs}\mathbb{E}\partial_r\partial_s\mathcal{R} 2 1 ∑ r , s Δ r s E ∂ r ∂ s R . Since
∂ r ∂ s R = R H r R H s R + R H s R H r R , \partial_r\partial_s\mathcal{R}=RH_rRH_s\mathcal{R}+RH_sRH_r\mathcal{R}, ∂ r ∂ s R = R H r R H s R + R H s R H r R ,
symmetry of Δ \Delta Δ combines the two terms into the covariance map in the preceding display. By (52) and ∥ R t ∥ ≤ η − 1 \lVert\mathcal{R}_t\rVert\leq\eta^{-1} ∥ R t ∥ ≤ η − 1 , the derivative has norm at most C N − 1 η − 3 CN^{-1}\eta^{-3} C N − 1 η − 3 . Integrating in t t t and taking normalized traces proves the lemma. Dominated convergence using the same resolvent bound gives continuity at t = 0 , 1 t=0,1 t = 0 , 1 . The argument uses no inverse covariance, so it includes the singular Gaussian endpoint ρ = 1 \rho=1 ρ = 1 . □
Transferring the count from the pencil to the polynomial
Choose a polynomial scale at which the comparison errors are dominated by the Gaussian bound. Set α = min ( ε , 1 ) \alpha=\min(\varepsilon,1) α = min ( ε , 1 ) and let μ ^ X , N \widehat{\mu}_{X,N} μ X , N be the eigenvalue law of H z ( X ) \mathcal{H}_z(X) H z ( X ) . The two comparison lemmas and (48) , together with E μ ^ X , N ( [ − η , η ] ) ≤ 2 η ℑ m X , N ( i η ) \mathbb{E}\widehat{\mu}_{X,N}([-\eta,\eta])\leq2\eta\Im m_{X,N}(i\eta) E μ X , N ([ − η , η ]) ≤ 2 η ℑ m X , N ( i η ) , give
E μ ^ X , N ( [ − η , η ] ) ≤ C ( η κ + N − α / 2 η − 2 − α + N − 1 η − 2 ) . \mathbb{E}\widehat{\mu}_{X,N}([-\eta,\eta])\leq C\left(\eta^\kappa+N^{-\alpha/2}\eta^{-2-\alpha}+N^{-1}\eta^{-2}\right). E μ X , N ([ − η , η ]) ≤ C ( η κ + N − α /2 η − 2 − α + N − 1 η − 2 ) .
Choose a 2 > 0 a_2>0 a 2 > 0 so that
a 2 < a 1 , a 2 ( κ + 2 + α ) < α / 2 , a 2 ( κ + 2 ) < 1. a_2<a_1,\qquad a_2(\kappa+2+\alpha)<\alpha/2,\qquad a_2(\kappa+2)<1. a 2 < a 1 , a 2 ( κ + 2 + α ) < α /2 , a 2 ( κ + 2 ) < 1.
For N − a 2 ≤ η ≤ 1 N^{-a_2}\leq\eta\leq1 N − a 2 ≤ η ≤ 1 , each error is then at most C η κ C\eta^\kappa C η κ .
To reach P − z I P-zI P − z I , the Gaussian argument used bounded elimination factors. Under the present moment assumptions, truncate these factors and bound the discarded rank in expectation, without an operator-norm assumption on X X X . Write
L z = A L Δ z A R , A L = V z − 1 , A R = W z − 1 , Δ z = diag ( P − z I , I ) . \mathcal{L}_z=A_L\Delta_zA_R,\qquad A_L=\mathcal{V}_z^{-1},\qquad A_R=\mathcal{W}_z^{-1},\qquad\Delta_z=\operatorname{diag}(P-zI,I). L z = A L Δ z A R , A L = V z − 1 , A R = W z − 1 , Δ z = diag ( P − z I , I ) .
The matrices A L , A R A_L,A_R A L , A R are fixed matrix-valued ∗ * ∗ -polynomials. Set their singular values above M M M to zero, obtaining A L , M , A R , M A_{L,M},A_{R,M} A L , M , A R , M of norm at most M M M . If r L , r R r_L,r_R r L , r R are the respective rank losses, then
rank ( L z − A L , M Δ z A R , M ) ≤ r L + r R , s j ( A L , M Δ z A R , M ) ≤ M 2 s j ( Δ z ) . \operatorname{rank}(\mathcal{L}_z-A_{L,M}\Delta_zA_{R,M})\leq r_L+r_R,\qquad s_j(A_{L,M}\Delta_zA_{R,M})\leq M^2s_j(\Delta_z). rank ( L z − A L , M Δ z A R , M ) ≤ r L + r R , s j ( A L , M Δ z A R , M ) ≤ M 2 s j ( Δ z ) .
For 0 < x < 1 0<x<1 0 < x < 1 , the identity block contributes nothing to the count of s j ( Δ z ) ≤ x s_j(\Delta_z)\leq\sqrt{x} s j ( Δ z ) ≤ x . Every singular value counted there is therefore counted in A L , M Δ z A R , M A_{L,M}\Delta_zA_{R,M} A L , M Δ z A R , M at threshold M 2 x M^2\sqrt{x} M 2 x . Rank interlacing with L z \mathcal{L}_z L z now gives, for N N z ( x ) = N − 1 # { j : s j ( P − z I ) 2 ≤ x } \mathcal{N}_N^z(x)=N^{-1}\#\{j:s_j(P-zI)^2\leq x\} N N z ( x ) = N − 1 # { j : s j ( P − z I ) 2 ≤ x } ,
N N z ( x ) ≤ h μ ^ X , N ( [ − M 2 x , M 2 x ] ) + r L + r R N , 0 < x < 1. (53) \mathcal{N}_N^z(x)\leq h\widehat{\mu}_{X,N}([-M^2\sqrt{x},M^2\sqrt{x}])+\frac{r_L+r_R}{N},\qquad0<x<1. \tag*{(53)} N N z ( x ) ≤ h μ X , N ([ − M 2 x , M 2 x ]) + N r L + r R , 0 < x < 1. (53)
The factor h h h follows because Hermitization counts each singular value twice and uses normalization 2 h N 2hN 2 h N .
By Lemma 7.1 , E ( r L + r R ) ≤ C N M − r ∗ \mathbb{E}(r_L+r_R)\leq CNM^{-r_*} E ( r L + r R ) ≤ C N M − r ∗ for some fixed r ∗ > 0 r_*>0 r ∗ > 0 . Choose M = x − 1 / 8 M=x^{-1/8} M = x − 1/8 . Then M 2 x = x 1 / 4 M^2\sqrt{x}=x^{1/4} M 2 x = x 1/4 , so the preceding pencil count and (53) yield
E N N z ( x ) ≤ C x ϑ , N − a ≤ x ≤ x ∗ , ϑ = min ( κ / 4 , r ∗ / 8 ) > 0 , (54) \mathbb{E}\mathcal{N}_N^z(x)\leq Cx^\vartheta,\qquad N^{-a}\leq x\leq x_*,\qquad\vartheta=\min(\kappa/4,r_*/8)>0, \tag*{(54)} E N N z ( x ) ≤ C x ϑ , N − a ≤ x ≤ x ∗ , ϑ = min ( κ /4 , r ∗ /8 ) > 0 , (54)
where any sufficiently small a < 4 a 2 a<4a_2 a < 4 a 2 is admissible. The count is unconditional, since the rank errors have been estimated in expectation.
Logarithmic tails and pointwise convergence of potentials
For assertion (ii), combine the small-value count with the least singular value cutoff to control the negative logarithmic tail. The positive tail needs only second moments and the rank bound. Let d = deg P ≥ 1 d=\deg P\ge1 d = deg P ≥ 1 . Lemma 7.1 at threshold x \sqrt{x} x gives E ν N z ( ( x , ∞ ) ) ≤ C x − 1 / d \mathbb{E}\nu_N^z((x,\infty))\le Cx^{-1/d} E ν N z (( x , ∞ )) ≤ C x − 1/ d . Integrating this estimate yields
sup N E ∫ ( T , ∞ ) log x ν N z ( d x ) ≤ C T − 1 / d log T + C ∫ T ∞ x − 1 − 1 / d d x ≤ C T − 1 / d ( log T + d ) , T ≥ 1. (55) \sup_N \mathbb{E}\int_{(T,\infty)}\log x\,\nu_N^z(\mathrm{d}x)
\le CT^{-1/d}\log T+C\int_T^\infty x^{-1-1/d}\,\mathrm{d}x
\le CT^{-1/d}(\log T+d),\qquad T\ge1. \tag*{(55)} N sup E ∫ ( T , ∞ ) log x ν N z ( d x ) ≤ C T − 1/ d log T + C ∫ T ∞ x − 1 − 1/ d d x ≤ C T − 1/ d ( log T + d ) , T ≥ 1. (55)
The count alone does not exclude a few extremely small singular values. For the lower tail, Corollary 1.3 gives a fixed B B B , sufficiently large that
F N = { s min ( P − z I ) ≥ N − B } , P ( F N c ) ⟶ 0. \mathcal{F}_N=\{s_{\min}(P-zI)\ge N^{-B}\},\qquad\mathbb{P}(\mathcal{F}_N^c)\longrightarrow0. F N = { s m i n ( P − z I ) ≥ N − B } , P ( F N c ) ⟶ 0.
Set x N = N − a x_N=N^{-a} x N = N − a with a a a from (54) , enlarging B B B so that 2 B > a 2B>a 2 B > a . On F N \mathcal{F}_N F N , every squared singular value is at least N − 2 B N^{-2B} N − 2 B , and hence
∫ ( 0 , x N ] ∣ log x ∣ ν N z ( d x ) ≤ 2 B log N N N N z ( x N ) . \int_{(0,x_N]}|\log x|\,\nu_N^z(\mathrm{d}x)\le2B\log N\,N\mathcal{N}_N^z(x_N). ∫ ( 0 , x N ] ∣ log x ∣ ν N z ( d x ) ≤ 2 B log N N N N z ( x N ) .
Its expectation after multiplication by 1 F N \mathbf{1}_{\mathcal{F}_N} 1 F N is at most C N − a ϑ log N = o ( 1 ) CN^{-a\vartheta}\log N=o(1) C N − a ϑ log N = o ( 1 ) . On the remaining interval, integration by parts gives, for x N ≤ δ ≤ min ( x ∗ , 1 ) x_N\le\delta\le\min(x_\ast,1) x N ≤ δ ≤ min ( x ∗ , 1 ) ,
E ∫ ( x N , δ ] ∣ log x ∣ ν N z ( d x ) ≤ ∣ log δ ∣ E N N z ( δ ) + ∫ x N δ E N N z ( x ) x d x ≤ C δ ϑ ( ∣ log δ ∣ + ϑ − 1 ) . \mathbb{E}\int_{(x_N,\delta]}|\log x|\,\nu_N^z(\mathrm{d}x)
\le|\log\delta|\,\mathbb{E}\mathcal{N}_N^z(\delta)+\int_{x_N}^{\delta}\frac{\mathbb{E}\mathcal{N}_N^z(x)}{x}\,\mathrm{d}x
\le C\delta^\vartheta(|\log\delta|+\vartheta^{-1}). E ∫ ( x N , δ ] ∣ log x ∣ ν N z ( d x ) ≤ ∣ log δ ∣ E N N z ( δ ) + ∫ x N δ x E N N z ( x ) d x ≤ C δ ϑ ( ∣ log δ ∣ + ϑ − 1 ) .
In particular, for every t > 0 t>0 t > 0 ,
lim δ ↓ 0 lim sup N → ∞ P { ∫ ( 0 , δ ] ∣ log x ∣ ν N z ( d x ) > t } = 0. \lim_{\delta\downarrow0}\limsup_{N\to\infty}\mathbb{P}\left\{\int_{(0,\delta]}|\log x|\,\nu_N^z(\mathrm{d}x)>t\right\}=0. δ ↓ 0 lim N → ∞ lim sup P { ∫ ( 0 , δ ] ∣ log x ∣ ν N z ( d x ) > t } = 0.
Indeed, Markov’s inequality on F N \mathcal{F}_N F N gives the bound after adding P ( F N c ) \mathbb{P}(\mathcal{F}_N^c) P ( F N c ) . The error q N q_N q N in the least-singular-value estimate appears only in this probability and acquires no factor log N \log N log N .
Assertion (i) now gives convergence for the logarithm truncated above and below. The bound (46) removes its lower truncation in the free law, whose upper tail vanishes by compact support. The preceding lower-tail estimate and (55) remove the two truncations for the random laws. Recalling that 1 2 ∫ log x d ν N z = N − 1 log ∣ det ( P − z I ) ∣ \frac{1}{2}\int\log x\,\mathrm{d}\nu_N^z=N^{-1}\log|\det(P-zI)| 2 1 ∫ log x d ν N z = N − 1 log ∣ det ( P − z I ) ∣ , we obtain
U N ( z ) : = N − 1 log ∣ det ( P ( X , X ∗ ) − z I ) ∣ → P U ( z ) : = τ ( log ∣ P ( c , c ∗ ) − z ∣ ) , z ≠ P ( 0 ) . (56) U_N(z):=N^{-1}\log|\det(P(X,X^\ast)-zI)|\xrightarrow{\mathbb{P}}U(z):=\tau(\log|P(c,c^\ast)-z|),\qquad z\ne P(0). \tag*{(56)} U N ( z ) := N − 1 log ∣ det ( P ( X , X ∗ ) − z I ) ∣ P U ( z ) := τ ( log ∣ P ( c , c ∗ ) − z ∣ ) , z = P ( 0 ) . (56)
Spatial integrability and Brown measure convergence
To pass to the Laplacian, we upgrade pointwise potential convergence to convergence after integrating the absolute difference over compact planar sets. A uniform spatial square-integral bound gives the required uniform integrability. We obtain it from the logarithmic kernel and eigenvalue product inequalities, starting with the consequence of (44) ,
sup N E 1 N ∑ j ( log + s j ( P ) ) 2 < ∞ . \sup_N\mathbb{E}\,\frac{1}{N}\sum_j(\log^+s_j(P))^2<\infty. N sup E N 1 j ∑ ( log + s j ( P ) ) 2 < ∞.
To apply the eigenvalue product inequalities, write a j = log + ∣ λ j ( P ) ∣ a_j=\log^{+}|\lambda_j(P)| a j = log + ∣ λ j ( P ) ∣ and b j = log + s j ( P ) b_j=\log^{+}s_j(P) b j = log + s j ( P ) in decreasing order. For a fixed k k k , let r ≤ k r\le k r ≤ k be the number of positive eigenvalue logarithms among the first k k k . Then
∑ j ≤ k a j = log ∏ j ≤ r ∣ λ j ( P ) ∣ ≤ log ∏ j ≤ r s j ( P ) ≤ ∑ j ≤ k b j . \sum_{j\le k}a_j=\log\prod_{j\le r}|\lambda_j(P)|\le\log\prod_{j\le r}s_j(P)\le\sum_{j\le k}b_j. j ≤ k ∑ a j = log j ≤ r ∏ ∣ λ j ( P ) ∣ ≤ log j ≤ r ∏ s j ( P ) ≤ j ≤ k ∑ b j .
Thus, for t ≥ 0 t\ge0 t ≥ 0 ,
∑ j ( a j − t ) + = max 0 ≤ k ≤ N ( ∑ j ≤ k a j − k t ) ≤ ∑ j ( b j − t ) + . \sum_j(a_j-t)_+=\max_{0\le k\le N}\left(\sum_{j\le k}a_j-kt\right)\le\sum_j(b_j-t)_+. j ∑ ( a j − t ) + = 0 ≤ k ≤ N max j ≤ k ∑ a j − k t ≤ j ∑ ( b j − t ) + .
Integrating against 2 t d t 2t\,dt 2 t d t proves ∑ j a j 2 ≤ ∑ j b j 2 \sum_j a_j^2\le\sum_j b_j^2 ∑ j a j 2 ≤ ∑ j b j 2 . In particular the eigenvalue measures are tight in probability, and their expected squared positive logarithms are uniformly bounded. For a compact planar set K K K , the elementary kernel estimate
∫ K ∣ log ∣ z − λ ∣ ∣ 2 d 2 z ≤ C K ( 1 + ( log + ∣ λ ∣ ) 2 ) \int_K|\log|z-\lambda||^2\,d^2z\le C_K(1+(\log^{+}|\lambda|)^2) ∫ K ∣ log ∣ z − λ ∣ ∣ 2 d 2 z ≤ C K ( 1 + ( log + ∣ λ ∣ ) 2 )
follows by integrating ∣ log ∣ z − λ ∣ ∣ 2 |\log|z-\lambda||^2 ∣ log ∣ z − λ ∣ ∣ 2 on ∣ z − λ ∣ ≤ 1 |z-\lambda|\le1 ∣ z − λ ∣ ≤ 1 and bounding the positive logarithm on the remaining part of K K K . Since U N ( z ) = N − 1 ∑ j log ∣ z − λ j ( P ) ∣ U_N(z)=N^{-1}\sum_j\log|z-\lambda_j(P)| U N ( z ) = N − 1 ∑ j log ∣ z − λ j ( P ) ∣ , Jensen’s inequality and the preceding eigenvalue bound give
sup N E ∫ K ∣ U N ( z ) ∣ 2 d 2 z < ∞ . (57) \sup_N\mathbb{E}\int_K|U_N(z)|^2\,d^2z<\infty. \tag*{(57)} N sup E ∫ K ∣ U N ( z ) ∣ 2 d 2 z < ∞. (57)
Realize all matrix sizes on a common product probability space. For any t > 0 t>0 t > 0 , (56) and bounded convergence in the planar variable give
∫ K P { ∣ U N ( z ) − U ( z ) ∣ > t } d 2 z ⟶ 0. \int_K\mathbb{P}\{|U_N(z)-U(z)|>t\}\,d^2z\longrightarrow0. ∫ K P { ∣ U N ( z ) − U ( z ) ∣ > t } d 2 z ⟶ 0.
The single excluded point P ( 0 ) P(0) P ( 0 ) has zero area. Hence U N U_N U N converges to U U U in measure on the finite product space. Take a subsequence converging almost everywhere there. Fatou’s lemma and (57) give U ∈ L 2 ( K ) U\in L^2(K) U ∈ L 2 ( K ) , and the same bound makes U N − U U_N-U U N − U uniformly integrable on the product space. It follows that
E ∫ K ∣ U N ( z ) − U ( z ) ∣ d 2 z ⟶ 0. \mathbb{E}\int_K|U_N(z)-U(z)|\,d^2z\longrightarrow0. E ∫ K ∣ U N ( z ) − U ( z ) ∣ d 2 z ⟶ 0.
Thus the potentials converge locally in expected spatial L 1 L^1 L 1 .
Finally Δ U N / ( 2 π ) \Delta U_N/(2\pi) Δ U N / ( 2 π ) is the empirical eigenvalue measure and Δ U / ( 2 π ) \Delta U/(2\pi) Δ U / ( 2 π ) is the Brown measure. For a smooth compactly supported ϕ \phi ϕ ,
∣ ∫ ϕ d μ P ( X , X ∗ ) − ∫ ϕ d μ P ( c , c ∗ ) ∣ ≤ ∥ Δ ϕ ∥ ∞ 2 π ∫ supp ϕ ∣ U N − U ∣ d 2 z , \left|\int\phi\,d\mu_{P(X,X^*)}-\int\phi\,d\mu_{P(c,c^*)}\right|
\le\frac{\|\Delta\phi\|_\infty}{2\pi}\int_{\operatorname{supp}\phi}|U_N-U|\,d^2z, ∫ ϕ d μ P ( X , X ∗ ) − ∫ ϕ d μ P ( c , c ∗ ) ≤ 2 π ∥Δ ϕ ∥ ∞ ∫ supp ϕ ∣ U N − U ∣ d 2 z ,
which tends to zero in probability. Eigenvalue tightness extends the convergence to bounded continuous tests. Constant P P P is immediate, and Theorem 1.1 follows.
Spectral transfer for Wigner inputs
For Theorem 1.4 , the limit is semicircular. The tail and spatial arguments apply once we compare Wigner atoms: an off-diagonal atom occupies conjugate positions, and the diagonal replacement must use only second moments. Off-diagonal laws have finite ( 2 + ε ) (2+\varepsilon) ( 2 + ε ) moments for some ε > 0 \varepsilon>0 ε > 0 ; diagonal laws need only second moments. The density assumptions enter through Theorem 1.5 . The entry laws may differ among input matrices. Define
ρ ℓ = E ζ ℓ , 12 2 , τ ℓ 2 = E d ℓ , 1 2 , α = min ( ε , 1 ) . \rho_\ell=\mathbb{E}\zeta_{\ell,12}^2,\qquad\tau_\ell^2=\mathbb{E}d_{\ell,1}^2,\qquad\alpha=\min(\varepsilon,1). ρ ℓ = E ζ ℓ , 12 2 , τ ℓ 2 = E d ℓ , 1 2 , α = min ( ε , 1 ) .
This includes real atoms (ρ ℓ = 1 \rho_\ell= 1 ρ ℓ = 1 ) and correlated planar atoms.
Fix a nonconstant scalar polynomial P P P in the Hermitian letters and z ≠ P ( 0 ) z \ne P(0) z = P ( 0 ) . Identify each letter with its adjoint in Lemma 2.1 and write the resulting pencil as
L z ( H ) = K z ⊗ I N + ∑ ℓ = 1 n C ℓ , z ⊗ H ℓ . L_z(H) = K_z \otimes I_N + \sum_{\ell=1}^{n} C_{\ell,z} \otimes H_\ell. L z ( H ) = K z ⊗ I N + ℓ = 1 ∑ n C ℓ , z ⊗ H ℓ .
If h h h is its coefficient size, its Hermitization is
H z ( H ) = D z ⊗ I N + ∑ ℓ B ℓ , z ⊗ H ℓ , D z = ( 0 K z K z ∗ 0 ) , B ℓ , z = ( 0 C ℓ , z C ℓ , z ∗ 0 ) . (58) \mathcal{H}_z(H) = D_z \otimes I_N + \sum_{\ell} B_{\ell,z} \otimes H_\ell,\qquad D_z = \begin{pmatrix} 0 & K_z \\ K_z^* & 0 \end{pmatrix},\quad B_{\ell,z} = \begin{pmatrix} 0 & C_{\ell,z} \\ C_{\ell,z}^* & 0 \end{pmatrix}. \tag*{(58)} H z ( H ) = D z ⊗ I N + ℓ ∑ B ℓ , z ⊗ H ℓ , D z = ( 0 K z ∗ K z 0 ) , B ℓ , z = ( 0 C ℓ , z ∗ C ℓ , z 0 ) . (58)
All B ℓ , z B_{\ell,z} B ℓ , z are Hermitian. For ℑ w > 0 \Im w > 0 ℑ w > 0 , define
m H , N ( w ) = E tr 2 h N ( H z ( H ) − w ) − 1 , η = ℑ w . m_{H,N}(w) = \mathbb{E}\operatorname{tr}_{2hN}\left(\mathcal{H}_z(H)-w\right)^{-1},\qquad\eta= \Im w. m H , N ( w ) = E tr 2 h N ( H z ( H ) − w ) − 1 , η = ℑ w .
The fixed shift z z z is suppressed in this notation.
Comparison of Wigner atoms. Let H m a t c h H_{\mathrm{match}} H match have independent centered Gaussian atoms matching the Wigner real covariances (off-diagonal absolute variance one, pseudovariance ρ ℓ \rho_\ell ρ ℓ , diagonal variance τ ℓ 2 \tau_\ell^2 τ ℓ 2 ). Let G 1 G U E , … , G n G U E G_1^{\mathrm{GUE}},\ldots,G_n^{\mathrm{GUE}} G 1 GUE , … , G n GUE be independent normalized GUE matrices, with both off-diagonal absolute variance and diagonal variance 1 / N 1/N 1/ N .
Lemma 7.4 (Wigner replacement and covariance comparison). For every ℑ w > 0 \Im w > 0 ℑ w > 0 ,
∣ m H , N ( w ) − m H m a t c h , N ( w ) ∣ ≤ C N − α / 2 η − 3 − α + C N − 1 η − 3 , (59) \left|m_{H,N}(w)-m_{H_{\mathrm{match}},N}(w)\right| \le CN^{-\alpha/2}\eta^{-3-\alpha}+CN^{-1}\eta^{-3}, \tag*{(59)} ∣ m H , N ( w ) − m H match , N ( w ) ∣ ≤ C N − α /2 η − 3 − α + C N − 1 η − 3 , (59) ∣ m H m a t c h , N ( w ) − m G G U E , N ( w ) ∣ ≤ C N − 1 / 2 η − 3 . (60) \left|m_{H_{\mathrm{match}},N}(w)-m_{\mathrm{GGUE},N}(w)\right| \le CN^{-1/2}\eta^{-3}. \tag*{(60)} ∣ m H match , N ( w ) − m GGUE , N ( w ) ∣ ≤ C N − 1/2 η − 3 . (60) Constants may depend on the fixed coefficients and atom laws. No density assumption is needed for these two estimates.
Proof. At a < b a<b a < b , keep the off-diagonal atom χ = v 1 + i v 2 \chi=v_1+iv_2 χ = v 1 + i v 2 together with its conjugate reverse entry. Its Hermitized perturbation is
A ( v ) = N − 1 / 2 B ℓ , z ⊗ ( χ E a b + χ ‾ E b a ) . A(v)=N^{-1/2}B_{\ell,z}\otimes\left(\chi E_{ab}+\overline{\chi}E_{ba}\right). A ( v ) = N − 1/2 B ℓ , z ⊗ ( χ E ab + χ E ba ) .
This is Hermitian, has bounded rank and norm at most C ∥ v ∥ / N C\lVert v\rVert/\sqrt{N} C ∥ v ∥ / N , and is independent of the background after conditioning on the other atoms. Apply the Taylor calculation of Lemma 7.2 to the two real coordinates together; its remainder is C N − 2 − α / 2 η − 3 − α ∥ v ∥ 2 + α CN^{-2-\alpha/2}\eta^{-3-\alpha}\lVert v\rVert^{2+\alpha} C N − 2 − α /2 η − 3 − α ∥ v ∥ 2 + α . Matching the entire real covariance, including its mixed real-imaginary entry, makes the quadratic expectations agree; centering does the same for the linear terms. Summing over O ( N 2 ) O(N^2) O ( N 2 ) atoms gives an error of C N − α / 2 η − 3 − α CN^{-\alpha/2}\eta^{-3-\alpha} C N − α /2 η − 3 − α .
For a diagonal atom the perturbation is N − 1 / 2 d B ℓ , z ⊗ E a a N^{-1/2}dB_{\ell,z}\otimes E_{aa} N − 1/2 d B ℓ , z ⊗ E aa . Expand only to first order. The first-order expectations agree by centering, and the second resolvent derivative bounds the expected remainder by C N − 2 η − 3 E ∣ d ∣ 2 CN^{-2}\eta^{-3}\mathbb{E}|d|^2 C N − 2 η − 3 E ∣ d ∣ 2 , together with the corresponding Gaussian term. Summing over the O ( N ) O(N) O ( N ) diagonal atoms gives C N − 1 η − 3 CN^{-1}\eta^{-3} C N − 1 η − 3 . This proves (59) without a moment above two on the diagonal.
For the Gaussian comparison, denote the centered Hermitized parts by Y m a t c h , Y G U E Y_{\mathrm{match}},Y_{\mathrm{GUE}} Y match , Y GUE , and define fixed linear maps K ∙ ( M ) = E [ Y ∙ M Y ∙ ] \mathcal{K}_{\bullet}(M)=\mathbb{E}[Y_{\bullet}MY_{\bullet}] K ∙ ( M ) = E [ Y ∙ M Y ∙ ] for deterministic M M M . The off-diagonal terms pairing an entry with its conjugate cancel because the absolute variances match. remaining terms are
( K m a t c h − K G U E ) ( M ) = 1 N ∑ ℓ , a < b [ ρ ℓ ( B ℓ , z ⊗ E a b ) M ( B ℓ , z ⊗ E a b ) + ρ ˉ ℓ ( B ℓ , z ⊗ E b a ) M ( B ℓ , z ⊗ E b a ) ] + 1 N ∑ ℓ , a ( τ ℓ 2 − 1 ) ( B ℓ , z ⊗ E a a ) M ( B ℓ , z ⊗ E a a ) . (61) (\mathcal{K}_{\mathrm{match}}-\mathcal{K}_{\mathrm{GUE}})(M)=\frac{1}{N}\sum_{\ell,a<b}\left[\rho_{\ell}(B_{\ell,z}\otimes E_{ab})M(B_{\ell,z}\otimes E_{ab})+\bar{\rho}_{\ell}(B_{\ell,z}\otimes E_{ba})M(B_{\ell,z}\otimes E_{ba})\right]
+\frac{1}{N}\sum_{\ell,a}(\tau_{\ell}^{2}-1)(B_{\ell,z}\otimes E_{aa})M(B_{\ell,z}\otimes E_{aa}). \tag*{(61)} ( K match − K GUE ) ( M ) = N 1 ℓ , a < b ∑ [ ρ ℓ ( B ℓ , z ⊗ E ab ) M ( B ℓ , z ⊗ E ab ) + ρ ˉ ℓ ( B ℓ , z ⊗ E ba ) M ( B ℓ , z ⊗ E ba ) ] + N 1 ℓ , a ∑ ( τ ℓ 2 − 1 ) ( B ℓ , z ⊗ E aa ) M ( B ℓ , z ⊗ E aa ) . (61)
To bound this map, consider one spatial N × N N \times N N × N block Q Q Q of M M M . The two off-diagonal sums act as
∑ a < b E a b Q E a b = tri + ( Q T ) , ∑ a < b E b a Q E b a = tri − ( Q T ) , \sum_{a<b}E_{ab}QE_{ab}=\operatorname{tri}_{+}(Q^{\mathsf{T}}),\qquad\sum_{a<b}E_{ba}QE_{ba}=\operatorname{tri}_{-}(Q^{\mathsf{T}}), a < b ∑ E ab Q E ab = tri + ( Q T ) , a < b ∑ E ba Q E ba = tri − ( Q T ) ,
where tri + \operatorname{tri}_{+} tri + and tri − \operatorname{tri}_{-} tri − denote strict upper and lower triangular truncation. Their weights are ρ ℓ \rho_{\ell} ρ ℓ and ρ ˉ ℓ \bar{\rho}_{\ell} ρ ˉ ℓ , so for nonreal pseudovariance the sum need not be a scalar multiple of the full transpose. We bound each part by
∥ tri ± ( Q T ) ∥ ≤ ∥ tri ± ( Q T ) ∥ H S ≤ ∥ Q ∥ H S ≤ N ∥ Q ∥ . \left\|\operatorname{tri}_{\pm}(Q^{\mathsf{T}})\right\|\leq\left\|\operatorname{tri}_{\pm}(Q^{\mathsf{T}})\right\|_{\mathrm{HS}}\leq\|Q\|_{\mathrm{HS}}\leq\sqrt{N}\|Q\|. tri ± ( Q T ) ≤ tri ± ( Q T ) HS ≤ ∥ Q ∥ HS ≤ N ∥ Q ∥.
The diagonal sum is diag ( Q 11 , … , Q N N ) \operatorname{diag}(Q_{11},\ldots,Q_{NN}) diag ( Q 11 , … , Q N N ) and has norm at most ∥ Q ∥ \|Q\| ∥ Q ∥ . Since the coefficient block dimensions are fixed, (61) gives
∥ ( K m a t c h − K G U E ) ( M ) ∥ ≤ C N − 1 / 2 ∥ M ∥ . (62) \|(\mathcal{K}_{\mathrm{match}}-\mathcal{K}_{\mathrm{GUE}})(M)\|\leq CN^{-1/2}\|M\|. \tag*{(62)} ∥ ( K match − K GUE ) ( M ) ∥ ≤ C N − 1/2 ∥ M ∥. (62)
Apply the Gaussian interpolation identity from Lemma 7.3 to independent Y m a t c h , Y G U E Y_{\mathrm{match}},Y_{\mathrm{GUE}} Y match , Y GUE . By (62) , its derivative has norm at most C N − 1 / 2 η − 3 CN^{-1/2}\eta^{-3} C N − 1/2 η − 3 . Integrating and taking the normalized trace gives (60) . The resolvent bound also gives continuity at the endpoints, including singular real Gaussian covariances. □
The polynomial count. We combine the Wigner comparison with the GUE reference bound and the expected rank loss from truncating the elimination factors.
Lemma 7.5 (Wigner mesoscopic singular-value count). Let
N H , N z ( x ) = N − 1 # { j : s j ( P ( H ) − z I N ) 2 ≤ x } . \mathcal{N}_{H,N}^{z}(x)=N^{-1}\#\{j:s_{j}(P(H)-zI_{N})^{2}\leq x\}. N H , N z ( x ) = N − 1 # { j : s j ( P ( H ) − z I N ) 2 ≤ x } . There are fixed a , ϑ , x ∗ , C > 0 a,\vartheta,x_{\ast},C>0 a , ϑ , x ∗ , C > 0 , depending on P P P , z z z and the atom laws, such that
E N H , N z ( x ) ≤ C x ϑ , N − a ≤ x ≤ x ∗ . (63) \mathbb{E}\mathcal{N}_{H,N}^{z}(x)\leq Cx^{\vartheta},\qquad N^{-a}\leq x\leq x_{\ast}. \tag*{(63)} E N H , N z ( x ) ≤ C x ϑ , N − a ≤ x ≤ x ∗ . (63) This estimate is unconditional.
Proof. The argument of § 7.2 applies to independent normalized GUE matrices. The squared singular-value limit is ν s c z \nu_{\mathrm{sc}}^{z} ν sc z , the law of Q z = ( P ( s ) − z ) ( P ( s ) − z ) ∗ Q_z=(P(s)-z)(P(s)-z)^{*} Q z = ( P ( s ) − z ) ( P ( s ) − z ) ∗ for free semicircular s 1 , … , s n s_1,\ldots,s_n s 1 , … , s n . The highest homogeneous term P d P d ∗ ≠ 0 P_dP_d^{*}\ne0 P d P d ∗ = 0 makes Q z Q_z Q z nonconstant, so [21 Theorem 1.1 and Corollary 1.2] and the Puiseux argument for (46) give
ν s c z ( [ 0 , x ] ) ≤ C x θ , 0 < x ≤ x ∗ , θ > 0. (64) \nu_{\mathrm{sc}}^{z}([0,x])\leq Cx^{\theta},\qquad0<x\leq x_{\ast},\qquad\theta>0. \tag*{(64)} ν sc z ([ 0 , x ]) ≤ C x θ , 0 < x ≤ x ∗ , θ > 0. (64)
In particular its negative logarithm is integrable. Applying [3 Lemma 5.5.4] directly to Q z Q_z Q z , choosing the cutoff as in (47) , gives
E N − 1 # { j : s j ( P ( G G U E ) − z I ) 2 ≤ x } ≤ C x θ , N − a 0 ≤ x ≤ 1 (65) \mathbb{E}N^{-1}\#\{j:s_j(P(G_{\mathrm{GUE}})-zI)^{2}\leq x\}\leq Cx^{\theta},\qquad N^{-a_0}\leq x\leq1 \tag*{(65)} E N − 1 # { j : s j ( P ( G GUE ) − z I ) 2 ≤ x } ≤ C x θ , N − a 0 ≤ x ≤ 1 (65)
for some a 0 > 0 a_0>0 a 0 > 0 . For GUE, the real and imaginary parts of u ∗ G G U E ∗ v u^{*}G_{\mathrm{GUE}}^{*}v u ∗ G GUE ∗ v have variance at most C / N C/N C / N for fixed complex unit vectors u , v u,v u , v . The net argument of § 7.2 gives P { max ℓ ∥ G ℓ G G U E ∥ > C 0 } ≤ C e − c N \mathbb{P}\{\max_{\ell}\lVert G_{\ell}^{\mathrm{GGUE}}\rVert>C_{0}\}\leq Ce^{-cN} P { max ℓ ∥ G ℓ GGUE ∥ > C 0 } ≤ C e − c N . On the complementary event the elimination factors are bounded, so the min–max argument and shell summation in (48) give
I m m G G U E , N ( i η ) ≤ C η κ − 1 , N − a 1 ≤ η ≤ 1 , (66) \mathop{\mathrm{Im}}m_{\mathrm{GGUE},N}(i\eta)\leq C\eta^{\kappa-1},\qquad N^{-a_{1}}\leq\eta\leq1, \tag*{(66)} Im m GGUE , N ( i η ) ≤ C η κ − 1 , N − a 1 ≤ η ≤ 1 , (66)
where 0 < κ < min ( 2 θ , 1 ) 0<\kappa<\min(2\theta,1) 0 < κ < min ( 2 θ , 1 ) and a 1 > 0 a_{1}>0 a 1 > 0 .
Let μ ^ H , N \widehat{\mu}_{H,N} μ H , N be the eigenvalue law of H z ( H ) H_{z}(H) H z ( H ) . Lemma 7.4, (66) , and the resolvent mass inequality give
E μ ^ H , N ( [ − η , η ] ) ≤ 2 η I m m H , N ( i η ) ≤ C { η κ + N − α / 2 η − 2 − α + N − 1 / 2 η − 2 } . \mathbb{E}\widehat{\mu}_{H,N}([-\eta,\eta])\leq2\eta\mathop{\mathrm{Im}}m_{H,N}(i\eta)\leq C\{\eta^{\kappa}+N^{-\alpha/2}\eta^{-2-\alpha}+N^{-1/2}\eta^{-2}\}. E μ H , N ([ − η , η ]) ≤ 2 η Im m H , N ( i η ) ≤ C { η κ + N − α /2 η − 2 − α + N − 1/2 η − 2 } .
The last term includes the smaller error from the diagonal atoms. Choose a 2 > 0 a_{2}>0 a 2 > 0 satisfying
a 2 < a 1 , a 2 ( κ + 2 + α ) < α / 2 , a 2 ( κ + 2 ) < 1 / 2. (67) \begin{aligned}
a_{2}&<a_{1},\\
a_{2}(\kappa+2+\alpha)&<\alpha/2,\\
a_{2}(\kappa+2)&<1/2. \tag*{(67)}
\end{aligned} a 2 a 2 ( κ + 2 + α ) a 2 ( κ + 2 ) < a 1 , < α /2 , < 1/2. (67)
Then the last bound is at most C η κ C\eta^{\kappa} C η κ for N − a 2 ≤ η ≤ 1 N^{-a_{2}}\leq\eta\leq1 N − a 2 ≤ η ≤ 1 .
The Wigner second moments satisfy E ∥ H ℓ ∥ H S 2 = N − 1 + τ ℓ 2 ≤ C N \mathbb{E}\lVert H_{\ell}\rVert_{\mathrm{HS}}^{2}=N-1+\tau_{\ell}^{2}\leq CN E ∥ H ℓ ∥ HS 2 = N − 1 + τ ℓ 2 ≤ C N . Lemma 7.1 thus applies to every fixed matrix-valued polynomial Q Q Q in the Hermitian letters, of degree at most D ≥ 1 D\geq1 D ≥ 1 , and gives
E # { j : s j ( Q ( H ) ) > M } ≤ C Q N M − 2 / D , M ≥ 1. (68) \mathbb{E}\#\{j:s_{j}(Q(H))>M\}\leq C_{Q}NM^{-2/D},\qquad M\geq1. \tag*{(68)} E # { j : s j ( Q ( H )) > M } ≤ C Q N M − 2/ D , M ≥ 1. (68)
Apply (53) and the argument of § 7.4 to the polynomial factors A L = V z − 1 A_{L}=\mathcal{V}_{z}^{-1} A L = V z − 1 and A R = W − 1 A_{R}=\mathcal{W}^{-1} A R = W − 1 : deleting their singular values above M M M has expected rank loss at most C N M − r ∗ CNM^{-r_{*}} C N M − r ∗ for some r ∗ > 0 r_{*}>0 r ∗ > 0 . Taking M = x − 1 / 8 M=x^{-1/8} M = x − 1/8 proves (63) with ϑ = min ( κ / 4 , r ∗ / 8 ) > 0 \vartheta=\min(\kappa/4,r_{*}/8)>0 ϑ = min ( κ /4 , r ∗ /8 ) > 0 and sufficiently small a < 4 a 2 a<4a_{2} a < 4 a 2 . The expected rank bounds make the count unconditional. □
The logarithmic limit. To complete the logarithmic integral, combine this count with weak convergence and the Wigner least singular value cutoff.
Lemma 7.6 (Finite-variance weak singular convergence for Wigner inputs). Under only the finite second moments of the Wigner model, the squared singular-value law of P ( H ) − z I P(H)-zI P ( H ) − z I converges weakly in probability to ν s c 2 \nu_{\mathrm{sc}}^{2} ν sc 2 .
Proof. For bounded laws, use the closed-walk expansion of § 7.1 , now indexed by unordered edges labeled by ℓ \ell ℓ . A surviving length-k k k pattern has v ≤ e + 1 ≤ 1 + k / 2 v\leq e+1\leq1+k/2 v ≤ e + 1 ≤ 1 + k /2 . Equality forces a tree traversed once in each direction, with each edge contributing E ∣ ζ ℓ , 12 ∣ 2 = 1 \mathbb{E}|\zeta_{\ell,12}|^{2}=1 E ∣ ζ ℓ , 12 ∣ 2 = 1 regardless of ρ ℓ \rho_{\ell} ρ ℓ . Loops (diagonal atoms) and parallel edges with different matrix labels lose a free index. Leaf removal gives the semicircular noncrossing pairings with matching matrix labels. The covariance count for connected walks is O ( N − 1 ) O(N^{-1}) O ( N − 1 ) ; compact support and moment determinacy give weak convergence in probability, as in § 7.1 .
For finite-variance laws, truncate and center the independent atoms at L L L , normalizing only the off-diagonal absolute variances to one. Reflecting the upper-triangular entries by conjugation gives a Hermitian coupling H ℓ , L H_{\ell,L} H ℓ , L . With generic atoms ζ ℓ , d ℓ \zeta_{\ell},d_{\ell} ζ ℓ , d ℓ and their truncated approximants on the right, it satisfies
E N − 1 ∥ H ℓ − H ℓ , L ∥ H S 2 = ( 1 − N − 1 ) E ∣ ζ ℓ − ζ ℓ , L ∣ 2 + N − 1 E ∣ d ℓ − d ℓ , L ∣ 2 . \mathbb{E}N^{-1}\lVert H_{\ell}-H_{\ell,L}\rVert_{\mathrm{HS}}^{2}=(1-N^{-1})\mathbb{E}|\zeta_{\ell}-\zeta_{\ell,L}|^{2}+N^{-1}\mathbb{E}|d_{\ell}-d_{\ell,L}|^{2}. E N − 1 ∥ H ℓ − H ℓ , L ∥ HS 2 = ( 1 − N − 1 ) E ∣ ζ ℓ − ζ ℓ , L ∣ 2 + N − 1 E ∣ d ℓ − d ℓ , L ∣ 2 .
The right side is at most δ L 2 → 0 \delta_{L}^{2}\to0 δ L 2 → 0 , uniformly in N , ℓ N,\ell N , ℓ . Diagonal variances do not affect the leading trees. Since truncation preserves Hermitian symmetry, we may apply (45) from § 7.1 . Choose T T T , then L L L , then let N → ∞ N\to\infty N → ∞ ; tightness from (68) proves the lemma. □
Proof of Theorem 1.4 . Lemma 7.6 gives weak singular-value convergence. Fix z ≠ P ( 0 ) z \ne P(0) z = P ( 0 ) . The pencil ( K z − 1 ⊗ I ) L z ( H ) (K_z^{-1} \otimes I)L_z(H) ( K z − 1 ⊗ I ) L z ( H ) has identity constant term. Theorem 1.5 and the inverse-block identity in Lemma 2.1 give a fixed B z < ∞ B_z < \infty B z < ∞ such that
F N , z = { s min ( P ( H ) − z I ) ≥ N − B z } , P ( F N , z c ) ⟶ 0. \mathcal{F}_{N,z}=\{s_{\min}(P(H)-zI)\ge N^{-B_z}\},\qquad\mathbb{P}(\mathcal{F}_{N,z}^{\mathrm{c}})\longrightarrow0. F N , z = { s m i n ( P ( H ) − z I ) ≥ N − B z } , P ( F N , z c ) ⟶ 0.
This cutoff uses only second moments, including on the diagonal.
Use the lower-tail argument of § 7.5 with (63) . Let x N = N − a x_N=N^{-a} x N = N − a and enlarge B z B_z B z so that 2 B z > a 2B_z>a 2 B z > a . For the squared singular-value law ν H , N # \nu_{H,N}^{\#} ν H , N # it gives
E 1 F N , z ∫ ( 0 , x N ] ∣ log x ∣ ν H , N # ( d x ) ≤ C N − a ϑ log N = o ( 1 ) , \mathbb{E}1_{\mathcal{F}_{N,z}}\int_{(0,x_N]}|\log x|\,\nu_{H,N}^{\#}(\mathrm{d}x)\le CN^{-a\vartheta}\log N=o(1), E 1 F N , z ∫ ( 0 , x N ] ∣ log x ∣ ν H , N # ( d x ) ≤ C N − a ϑ log N = o ( 1 ) ,
and, for x N ≤ δ ≤ min ( x ∗ , 1 ) x_N\le\delta\le\min(x_\ast,1) x N ≤ δ ≤ min ( x ∗ , 1 ) , the integral over ( x N , δ ] (x_N,\delta] ( x N , δ ] has expectation at most C δ ϑ ( ∣ log δ ∣ + ϑ − 1 ) C\delta^\vartheta(|\log\delta|+\vartheta^{-1}) C δ ϑ ( ∣ log δ ∣ + ϑ − 1 ) . Markov’s inequality on F N , z \mathcal{F}_{N,z} F N , z bounds the lower tail after adding the exceptional probability, without a factor log N \log N log N . The upper tail follows from (68) by (55) . Together with Lemma 7.6 and (64) , these estimates remove both logarithmic truncations and give
U H , N ( z ) : = N − 1 log ∣ det ( P ( H ) − z I ) ∣ → P U s c ( z ) : = τ ( log ∣ P ( s ) − z ∣ ) , z ≠ P ( 0 ) . U_{H,N}(z):=N^{-1}\log|\det(P(H)-zI)|\xrightarrow{\mathbb{P}}U_{\mathrm{sc}}(z):=\tau(\log|P(s)-z|),\qquad z\ne P(0). U H , N ( z ) := N − 1 log ∣ det ( P ( H ) − z I ) ∣ P U sc ( z ) := τ ( log ∣ P ( s ) − z ∣ ) , z = P ( 0 ) .
For spatial integrability, use (68) to bound the expected squared positive logarithms of singular values, as in § 7.6 . The eigenvalue product inequalities, planar kernel estimate, and uniform integrability give, for every compact K K K ,
E ∫ K ∣ U H , N ( z ) − U s c ( z ) ∣ d 2 z ⟶ 0. \mathbb{E}\int_K|U_{H,N}(z)-U_{\mathrm{sc}}(z)|\,\mathrm{d}^2z\longrightarrow0. E ∫ K ∣ U H , N ( z ) − U sc ( z ) ∣ d 2 z ⟶ 0.
Taking distributional Laplacians and using eigenvalue tightness gives weak convergence in probability to the Brown measure. The case of constant P P P is immediate, completing the proof. □