betaFeedback

WorksPhysics and Astronomy

Small-ball asymptotics of Gaussian matrix products: Open Fermi shells and the Grassmannian

Abstract

Let P=AN⋯A1P=A_N\cdots A_1 be a product of independent standard real or complex Gaussian matrices of arbitrary fixed sizes. For complex entries we determine the exact leading terms, of the form κ εC(log⁡1/ε)qκ\,\varepsilon^C(\log1/\varepsilon)^q, of the operator-norm and Frobenius small-ball probabilities of PP, together with the corresponding asymptotics of the Laplace transform of ∥P2∥F\|P^2\|_{F}. The exponent CC is the known diversity order, while the logarithmic power follows a filling rule: when fermions are placed on ladders of states determined by the sizes and the last occupied level has degeneracy μμ and contains hh particles, q=h(μ−h)q=h(μ-h). The determinant of this last shell expands in Plücker coordinates graded by the Schubert cells of the Grassmannian Gr(h,μ)\mathrm{Gr}(h,μ), with leading coefficients proportional to normalized degrees of Schubert varieties. For real entries we prove that the Frobenius small-ball probability has an exact positive leading term of order εC/2(log⁡1/ε)q\varepsilon^{C/2}(\log1/\varepsilon)^q and that the operator-norm probability has the same order, which gives a proof of the known real log canonical threshold and its multiplicity that needs no explicit resolution of singularities. We evaluate the real leading coefficient in closed form for every shell with at most one logarithm, a class that contains all products of two factors, and we conjecture its general form, which agrees with the classical constant of vector chains. Applications include the outage probability of cascaded multi-antenna channels at high signal-to-noise ratio and the Bayesian evidence of deep linear networks.

Introduction

Let F∈{R,C}F \in\{\mathbb{R},\mathbb{C}\} and let Aj∈Fdj×dj−1A_j \in\mathbb{F}^{d_j \times d_{j-1}}, j=1,…,Nj=1,\ldots,N, be independent matrices with independent standard Gaussian entries: N(0,1)N(0,1) if F=RF=\mathbb{R}, and density π−1e−∣z∣2\pi^{-1}e^{-|z|^2} on C\mathbb{C} if F=CF=\mathbb{C}. Put

P=AN⋯A1∈FdN×d0,n=min⁡0≤j≤Ndj.P=A_N\cdots A_1\in\mathbb{F}^{d_N\times d_0},\qquad n=\min_{0\leq j\leq N}d_j.

Almost surely PP has rank nn and we let x1,…,xn>0x_1,\ldots,x_n>0 be its nonzero squared singular values. For a fixed size vector d=(d0,…,dN)d=(d_0,\ldots,d_N) we study

Fop(ε)=P(∥P∥op2≤ε),FF(ε)=P(∥P∥F2≤ε),Z(t)=Ee−t∥P∥F2,(1)F_{\mathrm{op}}(\varepsilon)=\mathbb{P}\left(\lVert P\rVert_{\mathrm{op}}^2\leq\varepsilon\right),\qquad F_F(\varepsilon)=\mathbb{P}\left(\lVert P\rVert_F^2\leq\varepsilon\right),\qquad Z(t)=\mathbb{E}e^{-t\lVert P\rVert_F^2}, \tag*{(1)}

as ε↓0\varepsilon\downarrow0 and t→∞t \to\infty. Thus Fop(ε)F_{\mathrm{op}}(\varepsilon) is the probability that all nonzero squared singular values are at most ε\varepsilon, and FFF_F is the probability that their sum is. These are small-ball probabilities of the product around its most singular point P=0P = 0. They are finite-size quantities, complementary to the hard-edge gap probability, which concerns the smallest singular value, and to the large-nn and large-NN limits that dominate the literature on products [1, 2, 3, 4].

On the other hand, the Laplace transform Z(t)Z(t) is a common object behind several of the structures studied here. Up to the rescaling t↦t/2t \mapsto t/2, it is the classical partition function, at inverse temperature tt, of the energy H(A)=12∥AN⋯A1∥F2H(A) = \frac{1}{2}\lVert A_N \cdots A_1\rVert_F^2 with respect to the Gaussian reference measure, and for the zero-target deep linear networks recalled below this partition function is the Bayesian evidence. The limit t→∞t \to\infty is a zero-temperature limit that probes the geometry of the zero set {P=0}\{P = 0\} of the energy. Heuristically, the exponent is half the number of real directions transverse to this set, as in equipartition. In the same picture the logarithms come from continuous families of equally costly degenerations of the product that interpolate between competing components of {P=0}\{P = 0\}, and the scales along these families are organized by Schubert combinatorics, and we develop this picture in § 3.4. Our results show that these singular asymptotics admit a fermionic filling description and a Schubert-graded Grassmannian structure. We note that the fermionic picture is combinatorial: the energy of a filling is the codimension of a rank stratum of {P=0}\{P = 0\} (Proposition 2.9), and for complex entries the same filling arises from the determinantal structure of the product ensemble.

The same three quantities appear in three rather different fields. In wireless communication, products of Gaussian matrices model double- and multiple-scattering, keyhole and cascaded MIMO channels [5, 6, 7, 8]. The outage probability of optimal beamforming is FopF_{\mathrm{op}} evaluated at a multiple of the inverse signal-to-noise ratio, and that of orthogonal space-time block codes is FFF_F (§ 8.1). Yang and Belfiore determined the diversity order, that is, the exponent of the small-ball probability, for all sizes [7, 8]. For two factors the distribution of the largest eigenvalue is known in determinantal form [9, 10], but logarithmic corrections and coding gains have been extracted from it only in special cases, essentially when one node has a single antenna [9, 11, 12, 13], and for the closely related dual-hop relay channel Zhong et al. describe the general case as challenging [13]. For multi-layer scattering channels, Taricco studied the diversity–multiplexing tradeoff [14], and Wei, Zheng, Corander and Taricco approximated the outage of orthogonal space-time block codes in closed form by matching exact moments [15]. In singular learning theory, Z(t/2)Z(t/2) is the Bayesian evidence at sample size tt of a zero-target deep linear network with Gaussian prior. In Watanabe’s theory [16] its decay t−λ(log⁡t)m−1t^{-\lambda}(\log t)^{m-1} is governed by the real log canonical threshold λ\lambda, or learning coefficient, and by its multiplicity mm. For deep linear networks both were computed by Avogt through resolution of singularities [17, 18], and Pépin Lehalleu and Rimányi, using quiver representations, identified λ\lambda with half the codimension of the zero fibre of the multiplication map [19 Thm. 8.6]. The constant in front of t−λ(log⁡t)m−1t^{-\lambda}(\log t)^{m-1}, which enters the free energy at order one, is known for networks with one-dimensional output, whose evidence Hanin and Zlokapa computed exactly in terms of Meijer-GG functions [20]. Finally, in the geometry of matrix multiplication, the zero fibre Σd={P=0}\Sigma_d = \{P = 0\} and the ideal IdI_d generated by the entries of PP have been studied in [19, 21, 22]. Their codimension CC and the number θ\theta of top-dimensional components are invariant under permutations of dd, and E ∥P∥F−2z\mathbb{E}\,\lVert P\rVert_{\mathrm{F}}^{-2z} is the Gaussian average of the archimedean zeta function of IdI_d.

In § 8.2 we record the resulting learning coefficients, multiplicities and leading constants of deep linear networks, together with the exact integration over the last layer.

Our results are organized by a filling rule, which we explain in § 2. Deleting one entry equal to nn from dd and subtracting nn from the others gives offsets ν1,…,νN\nu_1,\ldots,\nu_N. The Mellin factors Γ(s+νj)\Gamma(s+\nu_j) of the product ensemble define NN ladders of one-particle states at energies νj,νj+1,…\nu_j,\nu_j+1,\ldots, which we fill with nn fermions. The highest occupied level, the Fermi level, has degeneracy μ\mu and occupation hh, and we put q=h(μ−h)q=h(\mu-h) (Figure 1). For complex entries we prove in Theorem A that

The filling rule for $d=(4,4,4,5,7)$

Figure 1. The filling rule for d=(4,4,4,5,7)d=(4,4,4,5,7), with offsets ν=(0,0,1,3)\nu=(0,0,1,3). Each ladder carries one state (circle) per energy, starting at its offset. The four fermions fill the two states at energy 0 (black) and any two of the three states at the Fermi level LF=1L_F=1 (grey), so that the shell is open with μ=3\mu=3, h=2h=2 and q=h(μ−h)=2q=h(\mu-h)=2. The (32)=3\binom{3}{2}=3 choices are the minimizers of Lemma 2.2 and the top components of Σd\Sigma_d. For the minimizer that fills the grey states on ladders 1 and 2, the first empty energies are f=(2,2,1,3)f=(2,2,1,3).

Fop(ε)∼κopεC(log⁡1/ε)q,FF(ε)∼κFεC(log⁡1/ε)q,Z(t)∼A t−C(log⁡t)q.F_{\mathrm{op}}(\varepsilon)\sim\kappa_{\mathrm{op}}\varepsilon^C(\log1/\varepsilon)^q,\qquad F_{\mathrm{F}}(\varepsilon)\sim\kappa_{\mathrm{F}}\varepsilon^C(\log1/\varepsilon)^q,\qquad Z(t)\sim\mathcal{A}\,t^{-C}(\log t)^q.

with explicit, generally distinct Gamma-product constants κop\kappa_{\mathrm{op}}, κF\kappa_{\mathrm{F}} and A\mathcal{A}, which Corollary 2.5 reformulates as lct⁡(Id)=C=codim⁡Σd\operatorname{lct}(I_d)=C=\operatorname{codim}\Sigma_d with multiplicity q+1q+1. The logarithms come from a shell determinant. In Theorem B we show that this determinant is a Plücker expansion whose terms are indexed by the Schubert cells of Gr⁡(h,μ)\operatorname{Gr}(h,\mu), with logarithmic degrees equal to the cell dimensions and leading coefficients proportional to normalized Schubert degrees. Consequently the number of Plücker terms is (μh)=χ(Gr⁡(h,μ))\binom{\mu}{h}=\chi(\operatorname{Gr}(h,\mu)), which by [19, 21, 22] is also the number θ\theta of top components of Σd\Sigma_d, the top degree is q=dim⁡Gr⁡(h,μ)q=\dim\operatorname{Gr}(h,\mu), and the shell factor of κop\kappa_{\mathrm{op}} and A\mathcal{A} is deg⁡Gr⁡(h,μ)/q!\deg\operatorname{Gr}(h,\mu)/q!.

For real entries we prove in Theorem C that FF(ε)∼κRεC/2(log⁡1/ε)qF_{\mathrm{F}}(\varepsilon)\sim\kappa_{\mathbb{R}}\varepsilon^{C/2}(\log1/\varepsilon)^q with κR>0\kappa_{\mathbb{R}}>0, and that FopF_{\mathrm{op}} has the same two-sided order, so that (C/2,q+1)(C/2,q+1) is the real log canonical threshold and multiplicity. Our proof is independent of Aoyagi’s theorem, on which the identification of the threshold in [19] rests, and of any explicit resolution of the multiplication map. It combines a positive-probability event built from the Bartlett decomposition with Theorem A, through a real–complex comparison of principalizations. The constant κR\kappa_{\mathbb{R}} is harder to reach, because the singular values of real products have no tractable joint law. We find it by first rewriting the complex constant as a product over pairs formed by an occupied state and the first empty state of another ladder (Proposition 2.8). Each factor in that product is a Gamma function Γ(fi−εp)\Gamma(f_i-\varepsilon_p) of the gap between an occupied energy εp\varepsilon_p and the first empty energy fif_i of another ladder. In Theorem D we show that for a closed shell, h=μh=\mu, and for a shell with a single logarithm, q=1q=1, the real constant contains the same pair product with every Gamma argument halved,

κR=2−NC/2Γ(C/2+1)∏jS1(ej)S1(n)∏(i,p)∗Γ(fi−εp2)∏j∏ej<u≤nΓ(νj+u2).\kappa_{\mathbb{R}}=\frac{2^{-NC/2}}{\Gamma(C/2+1)}\frac{\prod_j S_1(e_j)}{S_1(n)}\frac{\prod^{*}_{(i,p)}\Gamma\left(\frac{f_i-\varepsilon_p}{2}\right)}{\prod_j\prod_{e_j<u\le n}\Gamma\left(\frac{\nu_j+u}{2}\right)}.

where eje_j is the number of particles on ladder jj, S1(m)=∏t≤mΓ(t/2)S_1(m)=\prod_{t\le m}\Gamma(t/2), and ∏∗\prod^{*} runs over the ladders ii and the particles pp on other ladders with fi≠εpf_i\ne\varepsilon_p. The precise statement is in § 2.3. For a closed shell the proof evaluates a Morse–Bott integral along the unique top component of Σd\Sigma_d, in a slice adapted to the layered structure of PP in which the transversal Hessian is block diagonal. It then shows that no other divisor of a principalization contributes: over C\mathbb{C} this follows from Theorem A, and over R\mathbb{R} it follows from the real–complex comparison. For an open shell the same Morse–Bott integrand has exactly qq logarithmically divergent factors. Replacing each of them by the coefficient of its pole and multiplying by the shell factor reproduces Theorem A exactly, and on the real side it reproduces our closed form for two factors and the classical one for vector chains, the products in which some dj=1d_j=1. For q=1q=1 we prove this residue rule by an exact local normal form of the ideal of entries along the intersection of the two top components, and we conjecture that it gives κR\kappa_{\mathbb{R}} in general (Conjecture E).

References

  1. [1]Akemann, G., Ipsen, J.R., Kieburg, M.: Products of rectangular random matrices: singular values and progressive scattering. Physical Review E 88, 052118 (2013)DOI
  2. [2]Kuijlaars, A.B.J., Zhang, L.: Singular values of products of Ginibre random matrices, multiple orthogonal polynomials and hard edge scaling limits. Communications in Mathematical Physics 332, 759–781 (2014)DOI
  3. [3]Liu, D.-Z., Wang, D., Zhang, L.: Bulk and soft-edge universality for singular values of products of Ginibre random matrices. Annales de l’Institut Henri Poincaré, Probabilités et Statistiques 52, 1734–1762 (2016)arxiv.org/abs/1412.6777
  4. [4]Akemann, G., Burda, Z., Kieburg, M.: Universal distribution of Lyapunov exponents for products of Ginibre matrices. Journal of Physics A 47, 395202 (2014)DOI
  5. [5]Gesbert, D., Bölcskei, H., Gore, D.A., Paulraj, A.J.: Outdoor MIMO wireless channels: models and performance prediction. IEEE Transactions on Communications 50, 1926–1934 (2002)DOI
  6. [6]Müller, R.R.: A random matrix model of communication via antenna arrays. IEEE Transactions on Information Theory 48, 2495–2506 (2002)DOI
  7. [7]Yang, S., Belfiore, J.-C.: Diversity–multiplexing tradeoff of double scattering MIMO channels. IEEE Transactions on Information Theory 57, 2027–2034 (2011) arXiv:cs/0603124
  8. [8]Yang, S., Belfiore, J.-C.: Diversity of MIMO multihop relay channels. arXiv:0708.0386. See also: On the diversity of Rayleigh product channels, Proc. IEEE ISIT, pp. 1276–1280 (2007)arxiv.org/abs/0708.0386
  9. [9]Jin, S., McKay, M.R., Wong, K.-K., Gao, X.: Transmit beamforming in Rayleigh product MIMO channels: capacity and performance analysis. IEEE Transactions on Signal Processing 56, 5204–5221 (2008)DOI
  10. [10]Zhong, C., Jin, S., Wong, K.-K., McKay, M.R.: Outage analysis for optimal beamforming MIMO systems in multikeyhole channels. IEEE Transactions on Signal Processing 58, 1451–1456 (2010)DOI
  11. [11]Gong, Y., Letaief, K.B.: On the error probability of orthogonal space-time block codes over keyhole MIMO channels. IEEE Transactions on Wireless Communications 6, 3402–3409 (2007)DOI
  12. [12]Zhong, C.: Capacity and performance analysis of advanced multiple antenna communication systems. PhD thesis, University College London (2010)
  13. [13]Zhong, C., Ratnarajah, T., Jin, S., Wong, K.-K.: Performance analysis of optimal single stream beamforming in MIMO dual-hop AF systems. IEEE Journal on Selected Areas in Communications 30, 1415–1427 (2012) arXiv:1205.5614DOI
  14. [14]Tarrico, G.: Diversity–multiplexing tradeoff of multi-layer scattering MIMO channels. IEEE Wireless Communications Letters 9, 2045–2048 (2020)
  15. [15]Wei, L., Zheng, Z., Corander, J., Tarrico, G.: On the outage capacity of orthogonal space-time block codes over multi-cluster scattering MIMO channels. IEEE Transactions on Communications 63, 1700–1711 (2015) arXiv:1403.5571
  16. [16]Watanabe, S.: Algebraic Geometry and Statistical Learning Theory. Cambridge University Press, Cambridge (2009)
  17. [17]Aoyagi, M.: Consideration on the learning efficiency of multiple-layered neural networks with linear units. Neural Networks 172, 106132 (2024)DOI
  18. [18]Aoyagi, M.: Singular learning coefficients and efficiency in learning theory. arXiv:2501.12747 (2025)arxiv.org/abs/2501.12747
  19. [19]Pépin Lehalleur, S., Rimányi, R.: Geometry of fibers of the multiplication map of deep linear neural networks. arXiv:2411.19920 (2024)
  20. [20]Hanin, B., Zlokapa, A.: Bayesian interpolation with deep linear networks. Proceedings of the National Academy of Sciences 120, 2301345120 (2023) arXiv:2212.14457arxiv.org/abs/2212.14457
  21. [21]Koncki, J., Rimányi, R.: The main reasons for matrices multiplying to zero. Linear Algebra and its Applications 730, 587–602 (2026)
  22. [22]Terpai, T.: A note on numerical symmetry of the product-zero variety. arXiv:2607.09386 (2026)arxiv.org/abs/2607.09386
  23. [23]Ipsen, J.R., Kieburg, M.: Weak commutation relations and eigenvalue statistics for products of rectangular random matrices. Physical Review E 89, 032106 (2014)pubmed.ncbi.nlm.nih.gov/24730789
  24. [24]Hironaka, H.: Resolution of singularities of an algebraic variety over a field of characteristic zero. Annals of Mathematics 79, 109–326 (1964)
  25. [25]Włodarczyk, J.: Simple Hironaka resolution in characteristic zero. Journal of the American Mathematical Society 18, 779–822 (2005)
  26. [26]Kollár, J.: Lectures on Resolution of Singularities. Annals of Mathematics Studies, vol. 166. Princeton University Press, Princeton, NJ (2007)DOI
  27. [27]Aoyagi, M., Watanabe, S.: Stochastic complexities of reduced rank regression in Bayesian estimation. Neural Networks 18, 924–933 (2005)pubmed.ncbi.nlm.nih.gov/15993036
  28. [28]Forrester, P.J., Warnaar, S.O.: The importance of the Selberg integral. Bulletin of the American Mathematical Society 45, 489–534 (2008)arxiv.org/abs/0710.3981
  29. [29]Springer, M.D., Thompson, W.E.: The distribution of products of beta, gamma and Gaussian random variables. SIAM Journal on Applied Mathematics 18, 721–737 (1970)DOI
  30. [30]Fulton, W.: Young Tableaux. London Mathematical Society Student Texts, vol. 35. Cambridge University Press, Cambridge (1997)
  31. [31]Stanley, R.P.: Two poset polytopes. Discrete and Computational Geometry 1, 9–23 (1986)DOI
  32. [32]Shewchuk, J.R., Bhattacharya, S.: The geometry of the set of equivalent linear neural networks. arXiv:2404.14855 (2024)arxiv.org/abs/2404.14855
  33. [33]Trager, M., Kohn, K., Bruna, J.: Pure and spurious critical points: a geometric study of linear networks. In: International Conference on Learning Representations (ICLR) (2020)arxiv.org/abs/1910.01671
  34. [34]Shin, H., Win, M.Z.: MIMO diversity in the presence of double scattering. IEEE Transactions on Information Theory 54, 2976–2996 (2008)arxiv.org/abs/cs/0511028
  35. [35]Feller, W.: An Introduction to Probability Theory and Its Applications vol. II, 2nd edn. Wiley, New York (1971)
  36. [36]Atiyah, M.F.: Resolution of singularities and division of distributions. Communications on Pure and Applied Mathematics 23, 145–150 (1970)DOI
  37. [37]Bernstein, I.N., Gelfand, S.I.: Meromorphy of the function P^λ. Functional Analysis and Its Applications 3, 68–69 (1969)
  38. [38]Mustaţă, M.: IMPANGA lecture notes on log canonical thresholds. In: Contributions to Algebraic Geometry, pp. 407–442. European Mathematical Society, Zürich (2012)arxiv.org/abs/1107.2676
  39. [39]Aitken, A.C.: The monomial expansion of determinantal symmetric functions. Proceedings of the Royal Society of Edinburgh, Section A 61, 300–310 (1943)DOI
  40. [40]Stanley, R.P.: Enumerative Combinatorics vol. 2. Cambridge University Press, Cambridge (1999)
  41. [41]Eisenbud, D., Harris, J.: Divisors on general curves and cuspidal rational curves. Inventiones Mathematicae 74, 371–418 (1983)DOI
  42. [42]Purbhoo, K.: Jeu de taquin and a monodromy problem for Wronskians of polynomials. Advances in Mathematics 224, 827–862 (2010) arXiv:0902.1321arxiv.org/abs/0902.1321
  43. [43]Muirhead, R.J.: Aspects of Multivariate Statistical Theory. Wiley, New York (1982)
  44. [44]Newman, C.M.: The distribution of Lyapunov exponents: exact results for random matrices. Communications in Mathematical Physics 103, 121–126 (1986)DOI
  45. [45]Forrester, P.J.: Lyapunov exponents for products of complex Gaussian random matrices. Journal of Statistical Physics 151, 796–808 (2013) arXiv:1206.2001DOI
  46. [46]Saito, M.: On real log canonical thresholds. arXiv:0707.2308 (2007)arxiv.org/abs/0707.2308
  47. [47]Kosta, D., Windisch, D.: Classification of real hyperplane singularities by real log canonical thresholds. arXiv:2411.13392 (2024)arxiv.org/abs/2411.13392
  48. [48]Kieburg, M., Kuijlaars, A.B.J., Stivigny, D.: Singular value statistics of matrix products with truncated unitary matrices. International Mathematics Research Notices 2016, 3392–3424 (2016) arXiv:1501.03910DOI

Paper details

vlnc:W_qh87hdrq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 2

downloads · 0

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

  • Physics and Astronomy

GPT-6 Astra, Claude Opus 5.5

View statement
Looking for related papers…
Contents