{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,6]],"date-time":"2026-10-06T21:51:34Z","timestamp":1791323494981,"version":"4.3.1"},"reference-count":69,"publisher":"Institute of Mathematical Statistics","issue":"4","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["Ann. Statist."],"published-print":{"date-parts":[[2026,8,1]]},"abstract":"We study the local geometry of empirical risks in high dimensions via the spectral theory of their Hessian and information matrices. We focus on settings where the data, (Y\u2113)\u2113=1n\u2208Rd, are i.i.d. draws of a k-Gaussian mixture model, and the loss depends on the projection of the data into a fixed number of vectors, namely x\u22a4Y, where x\u2208Rd\u00d7C are the parameters, and C need not equal k. This setting captures a broad class of problems such as classification by one and two-layer networks and regression on multi-index models. We provide exact formulas for the limits of the empirical spectral distribution and outlier eigenvalues and eigenvectors of such matrices in the proportional asymptotics limit, where the number of samples and dimension n,d\u2192\u221e and n\/d=\u03d5\u2208(0,\u221e). These limits depend on the parameters x only through the summary statistic of the (C+k)\u00d7(C+k) Gram matrix of the parameters and class means, G=(x,\u03bc)\u22a4(x,\u03bc). It is known that under general conditions, when x is trained by online stochastic gradient descent, the evolution of these same summary statistics along training converges to the solution of an autonomous system of ODEs, called the effective dynamics. This enables us to connect the training dynamics to the spectral theory of these matrices generated with test data. We demonstrate our general results by analyzing the effective spectrum along the effective dynamics in the case of multiclass logistic regression. In this setting, the empirical Hessian and information matrices have substantially different spectra, each with their own static and dynamical spectral transitions.<\/jats:p>","DOI":"10.1214\/26-aos2621","type":"journal-article","created":{"date-parts":[[2026,8,18]],"date-time":"2026-08-18T08:11:25Z","timestamp":1787040685000},"page":"1682-1711","source":"Crossref","is-referenced-by-count":0,"title":["Local geometry of high-dimensional mixture models: Effective spectral theory and dynamical transitions"],"prefix":"10.1214","volume":"54","author":[{"given":"G\u00e9rard","family":"Ben Arous","sequence":"first","affiliation":[{"name":"Courant Institute, New York University"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Reza","family":"Gheissari","sequence":"additional","affiliation":[{"name":"Department of Mathematics, Northwestern University"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Jiaoyang","family":"Huang","sequence":"additional","affiliation":[{"name":"Department of Statistics and Data Science, University of Pennsylvania"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Aukosh","family":"Jagannath","sequence":"additional","affiliation":[{"name":"Department of Statistics and Actuarial Science, Department of Applied Mathematics, and Cheriton School of Computer Science, University of Waterloo"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"108","reference":[{"key":"15","doi-asserted-by":"publisher","unstructured":"Bishop, C. M.<\/small> (2006). Pattern Recognition and Machine Learning<\/i>. Information Science and Statistics<\/i>. Springer, New York.","DOI":"10.1007\/978-0-387-45528-0"},{"key":"62","doi-asserted-by":"publisher","unstructured":"SILVERSTEIN, J. W.<\/small> and BAI, Z. D.<\/small> (1995). On the empirical distribution of eigenvalues of a class of large-dimensional random matrices. J. Multivariate Anal.<\/i> 54<\/b> 175\u2013192.","DOI":"10.1006\/jmva.1995.1051"},{"key":"2","doi-asserted-by":"publisher","unstructured":"AMARI, S.<\/small> (2016). Information Geometry and Its Applications<\/i>. Applied Mathematical Sciences<\/i> 194<\/b>. Springer, Tokyo.","DOI":"10.1007\/978-4-431-55978-8"},{"key":"16","doi-asserted-by":"publisher","unstructured":"Cand\u00e8s, E. J.<\/small>, Li, X.<\/small> and Soltanolkotabi, M.<\/small> (2015). Phase retrieval via Wirtinger flow: Theory and algorithms. IEEE Trans. Inf. Theory<\/i> 61<\/b> 1985\u20132007.","DOI":"10.1109\/TIT.2015.2399924"},{"key":"24","doi-asserted-by":"publisher","unstructured":"Derrida, B.<\/small> (1981). Random-energy model: An exactly solvable model of disordered systems. Phys. Rev. B<\/i> (3) 24<\/b> 2613\u20132626.","DOI":"10.1103\/physrevb.24.2613"},{"key":"9","doi-asserted-by":"publisher","unstructured":"BARTLETT, P. L.<\/small>, MONTANARI, A.<\/small> and RAKHLIN, A.<\/small> (2021). Deep learning: A statistical viewpoint. Acta Numer.<\/i> 30<\/b> 87\u2013201.","DOI":"10.1017\/S0962492921000027"},{"key":"17","doi-asserted-by":"publisher","unstructured":"CHEN, Y.<\/small>, CHI, Y.<\/small>, FAN, J.<\/small> and MA, C.<\/small> (2019). Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval. Math. Program.<\/i> 176<\/b> 5\u201337.","DOI":"10.1007\/s10107-019-01363-6"},{"key":"64","doi-asserted-by":"publisher","unstructured":"TAN, Y. S.<\/small> and VERSHYNIN, R.<\/small> (2019). Phase retrieval via randomized Kaczmarz: Theoretical guarantees. Inf. Inference<\/i> 8<\/b> 97\u2013123.","DOI":"10.1093\/imaiai\/iay005"},{"key":"6","doi-asserted-by":"publisher","unstructured":"Bai, Z.<\/small> and Silverstein, J. W.<\/small> (2010). Spectral Analysis of Large Dimensional Random Matrices<\/i>, 2nd ed. Springer Series in Statistics<\/i>. Springer, New York.","DOI":"10.1007\/978-1-4419-0661-8"},{"key":"8","doi-asserted-by":"publisher","unstructured":"Baik, J.<\/small>, Ben Arous, G.<\/small> and P\u00e9ch\u00e9, S.<\/small> (2005). Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices. Ann. Probab.<\/i> 33<\/b> 1643\u20131697.","DOI":"10.1214\/009117905000000233"},{"key":"34","doi-asserted-by":"publisher","unstructured":"Knowles, A.<\/small> and Yin, J.<\/small> (2017). Anisotropic local laws for random matrices. Probab. Theory Related Fields<\/i> 169<\/b> 257\u2013352.","DOI":"10.1007\/s00440-016-0730-4"},{"key":"7","doi-asserted-by":"publisher","unstructured":"Bai, Z. D.<\/small> and Silverstein, J. W.<\/small> (1998). No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices. Ann. Probab.<\/i> 26<\/b> 316\u2013345.","DOI":"10.1214\/aop\/1022855421"},{"key":"31","doi-asserted-by":"publisher","unstructured":"Hastie, T.<\/small>, Tibshirani, R.<\/small> and Friedman, J.<\/small> (2009). The Elements of Statistical Learning<\/i>: Data Mining, Inference, and Prediction<\/i>, 2nd ed. Springer Series in Statistics<\/i>. Springer, New York.","DOI":"10.1007\/978-0-387-84858-7"},{"key":"18","unstructured":"CHIZAT, L.<\/small>, OYALLON, E.<\/small> and BACH, F.<\/small> (2019). On lazy training in differentiable programming. Adv. Neural Inf. Process. Syst.<\/i> 32<\/b>."},{"key":"39","doi-asserted-by":"publisher","unstructured":"LU, Y. M.<\/small> and LI, G.<\/small> (2020). Phase transitions of spectral initialization for high-dimensional non-convex estimation. Inf. Inference<\/i> 9<\/b> 507\u2013541.","DOI":"10.1093\/imaiai\/iaz020"},{"key":"46","unstructured":"MONDELLI, M.<\/small> and MONTANARI, A.<\/small> (2018). Fundamental limits of weak recovery with applications to phase retrieval. In Conference on Learning Theory<\/i> 1445\u20131450. PMLR."},{"key":"13","doi-asserted-by":"publisher","unstructured":"BEN AROUS, G.<\/small>, GHEISSARI, R.<\/small> and JAGANNATH, A.<\/small> (2024). High-dimensional limit theorems for SGD: Effective dynamics and critical scaling. Comm. Pure Appl. Math.<\/i> 77<\/b> 2030\u20132080.","DOI":"10.1002\/cpa.22169"},{"key":"1","unstructured":"ABBE, E.<\/small>, ADSERA, E. B.<\/small> and MISIAKIEWICZ, T.<\/small> (2023). SGD learning on neural networks: Leap complexity and saddle-to-saddle dynamics. In The Thirty Sixth Annual Conference on Learning Theory<\/i> 2552\u20132623. PMLR."},{"key":"4","doi-asserted-by":"crossref","unstructured":"ASGARI, K.<\/small>, MONTANARI, A.<\/small> and SAEED, B.<\/small> (2025). Local minima of the empirical risk in high dimension: General theorems and convex examples.","DOI":"10.1214\/26-AOS2652"},{"key":"5","unstructured":"AUBIN, B.<\/small>, LOUREIRO, B. V.<\/small>, BAKER, A.<\/small>, KRZAKALA, F.<\/small> and ZDEBOROV\u00c1, L.<\/small> (2019). Exact asymptotics for phase retrieval and compressed sensing with random generative priors. ArXiv. Available at arXiv:1912.02008<\/a>."},{"key":"10","unstructured":"BEN AROUS, G.<\/small>, GHEISSARI, R.<\/small>, HUANG, J.<\/small> and JAGANNATH, A.<\/small> (2024). High-dimensional SGD aligns with emerging outlier eigenspaces. In The Twelfth International Conference on Learning Representations<\/i>."},{"key":"11","doi-asserted-by":"crossref","unstructured":"BEN AROUS, G.<\/small>, GHEISSARI, R.<\/small>, HUANG, J.<\/small> and JAGANNATH, A.<\/small> (2026). Supplement to \u201cLocal geometry of high-dimensional mixture models: Effective spectral theory and dynamical transitions.\u201d https:\/\/doi.org\/10.1214\/26-AOS2621SUPP<\/a>","DOI":"10.1214\/26-AOS2621SUPP"},{"key":"12","unstructured":"BEN AROUS, G.<\/small>, GHEISSARI, R.<\/small> and JAGANNATH, A.<\/small> (2021). Online stochastic gradient descent on non-convex losses from high-dimensional inference. J. Mach. Learn. Res.<\/i> 22<\/b> 106."},{"key":"14","doi-asserted-by":"publisher","unstructured":"BIETTI, A.<\/small>, BRUNA, J.<\/small> and PILLAUD-VIVIEN, L.<\/small> (2025). On learning Gaussian multi-index models with gradient flow part I: General properties and two-timescale learning. Comm. Pure Appl. Math.<\/i> 78<\/b> 2354\u20132435.","DOI":"10.1002\/cpa.70006"},{"key":"20","doi-asserted-by":"publisher","unstructured":"COLLINS-WOODFIN, E.<\/small>, PAQUETTE, C.<\/small>, PAQUETTE, E.<\/small> and SEROUSSI, I.<\/small> (2024). Hitting the high-dimensional notes: An ODE for SGD learning dynamics on GLMs and multi-index models. Inf. Inference<\/i> 13<\/b> iaae028.","DOI":"10.1093\/imaiai\/iaae028"},{"key":"22","doi-asserted-by":"crossref","unstructured":"DAMIAN, A.<\/small>, NICHANI, E.<\/small>, GE, R.<\/small> and LEE, J. D.<\/small> (2023). Smoothing the landscape boosts the signal for SGD: Optimal sample complexity for learning single index models. In Advances in Neural Information Processing Systems<\/i> (A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt and S. Levine, eds.) 36<\/b> 752\u2013784. Curran Associates, Red Hook.","DOI":"10.52202\/075280-0035"},{"key":"23","doi-asserted-by":"crossref","unstructured":"DEFILIPPIS, L.<\/small>, DANDI, Y.<\/small>, MERGNY, P.<\/small>, KRZAKALA, F.<\/small> and LOUREIRO, B.<\/small> (2025). Optimal Spectral Transitions in High-Dimensional Multi-Index Models<\/i>.","DOI":"10.52202\/085713-5818"},{"key":"25","unstructured":"DUDEJA, R.<\/small> and HSU, D.<\/small> (2018). Learning single-index models in Gaussian space. In Proceedings of the<\/i> 31st Conference on Learning Theory<\/i>. Proceedings of Machine Learning Research<\/i> 75<\/b> 1887\u20131930. PMLR."},{"key":"26","unstructured":"FAN, Z.<\/small> and WANG, Z.<\/small> (2020). Spectra of the conjugate kernel and neural tangent kernel for linear-width neural networks. In Proceedings of the<\/i> 34th International Conference on Neural Information Processing Systems. NIPS\u2019<\/i>20 Curran Associates, Red, Hook, NY, USA."},{"key":"27","unstructured":"GARROD, C.<\/small> and KEATING, J. P.<\/small> (2024). Unifying Low Dimensional Observations in Deep Learning Through the Deep Linear Unconstrained Feature Model."},{"key":"29","doi-asserted-by":"crossref","unstructured":"GOLDT, S.<\/small>, ADVANI, M.<\/small>, SAXE, A. M.<\/small>, KRZAKALA, F.<\/small> and ZDEBOROV\u00c1, L.<\/small> (2019). Dynamics of stochastic gradient descent for two-layer neural networks in the teacher-student setup. Adv. Neural Inf. Process. Syst.<\/i> 32<\/b>.","DOI":"10.1088\/1742-5468\/abc61e"},{"key":"33","unstructured":"JEONG, H.<\/small> and G\u00dcNT\u00dcRK, C. S.<\/small> (2017). Convergence of the randomized Kaczmarz method for phase retrieval. ArXiv. Available at arXiv:1706.10291<\/a>."},{"key":"35","unstructured":"KOVACEVIC, F.<\/small>, ZHANG, Y.<\/small> and MONDELLI, M.<\/small> (2025). Spectral Estimators for Multi-Index Models: Precise Asymptotics and Optimal Weak Recovery."},{"key":"36","doi-asserted-by":"publisher","unstructured":"LECUN, Y. A.<\/small>, BOTTOU, L.<\/small>, ORR, G. B.<\/small> and M\u00dcLLER, K.-R.<\/small> (2012). Efficient BackProp. In Neural Networks<\/i>: Tricks of the Trade<\/i> 9\u201348 Second ed. Springer, Berlin.","DOI":"10.1007\/978-3-642-35289-8_3"},{"key":"37","doi-asserted-by":"publisher","unstructured":"LI, X.<\/small>, GU, Q.<\/small>, ZHOU, Y.<\/small>, CHEN, T.<\/small> and BANERJEE, A.<\/small> (2020). Hessian based analysis of SGD for deep nets: Dynamics and generalization. In Proceedings of the<\/i> 2020 SIAM International Conference on Data Mining<\/i> (SDM<\/i>) 190\u2013198.","DOI":"10.1137\/1.9781611976236.22"},{"key":"40","doi-asserted-by":"publisher","unstructured":"LUO, W.<\/small>, ALGHAMDI, W.<\/small> and LU, Y. M.<\/small> (2019). Optimal spectral initialization for signal recovery with applications to phase retrieval. IEEE Trans. Signal Process.<\/i> 67<\/b> 2347\u20132456.","DOI":"10.1109\/TSP.2019.2904918"},{"key":"42","unstructured":"MARCHENKO, V. A.<\/small> and PASTUR, L. A.<\/small> (1967). Distribution of eigenvalues for some sets of random matrices. Mat. Sb.<\/i> 114<\/b> 507\u2013536."},{"key":"43","doi-asserted-by":"publisher","unstructured":"MARRIOTT, P.<\/small> (2002). On the local geometry of mixture models. Biometrika<\/i> 89<\/b> 77\u201393.","DOI":"10.1093\/biomet\/89.1.77"},{"key":"45","unstructured":"MINSKY, M.<\/small> and PAPERT, S.<\/small> (1969). An introduction to computational geometry. Camb. Tiass<\/i>., HIT<\/i> 479<\/b> 480."},{"key":"48","doi-asserted-by":"crossref","unstructured":"MOUSAVI-HOSSEINI, A.<\/small>, WU, D.<\/small>, SUZUKI, T.<\/small> and ERDOGDU, M. A.<\/small> (2023). Gradient-based feature learning under structured data. In Advances in Neural Information Processing Systems<\/i> (A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt and S. Levine, eds.) 36<\/b> 71449\u201371485. Curran Associates, Red Hook.","DOI":"10.52202\/075280-3129"},{"key":"49","unstructured":"NIELSEN, F.<\/small>, RAO, A. SS.<\/small> and RAO, C. R.<\/small> (2022). Geometry and Statistics 46, San Diego Academic Press."},{"key":"50","unstructured":"OYMAK, S.<\/small> and SOLTANOLKOTABI, M.<\/small> (2019). Overparameterized nonlinear learning: Gradient descent takes the shortest path? In International Conference on Machine Learning<\/i> 4951\u20134960. PMLR."},{"key":"52","unstructured":"PAPYAN, V.<\/small> (2020). Traces of class\/cross-class structure pervade deep learning spectra. J. Mach. Learn. Res.<\/i> 21<\/b> 252."},{"key":"56","unstructured":"PENNINGTON, J.<\/small> and WORAH, P.<\/small> (2018). The spectrum of the Fisher information matrix of a single-hidden-layer neural network. In Advances in Neural Information Processing Systems<\/i> (S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi and R. Garnett, eds.). Curran Associates<\/i> 31<\/b>. Curran Associates, Red Hook."},{"key":"60","unstructured":"SAGUN, L.<\/small>, BOTTOU, L.<\/small> and LECUN, Y.<\/small> (2017). Eigenvalues of the Hessian in Deep Learning<\/i>: Singularity and Beyond<\/i>."},{"key":"63","unstructured":"SIMSEK, B.<\/small>, BENDJEDDOU, A.<\/small> and HSU, D.<\/small> (2024). Learning Gaussian Multi-Index Models with Gradient Flow<\/i>: Time Complexity and Directional Convergence<\/i>."},{"key":"65","unstructured":"TAN, Y. S.<\/small> and VERSHYNIN, R.<\/small> (2023). Online stochastic gradient descent with arbitrary initialization solves non-smooth, non-convex phase retrieval. J. Mach. Learn. Res.<\/i> 24<\/b> 58."},{"key":"66","doi-asserted-by":"publisher","unstructured":"VEIGA, R.<\/small>, STEPHAN, L.<\/small>, LOUREIRO, B.<\/small>, KRZAKALA, F.<\/small> and ZDEBOROV\u00c1, L.<\/small> (2023). Phase diagram of stochastic gradient descent in high-dimensional two-layer neural networks. J. Stat. Mech. Theory Exp.<\/i> 11<\/b> 114008.","DOI":"10.1088\/1742-5468\/ad01b1"},{"key":"68","unstructured":"ZHANG, H.<\/small> and LIANG, Y.<\/small> (2016). Reshaped Wirtinger flow for solving quadratic system of equations. In Advances in Neural Information Processing Systems<\/i> 29 (D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon and R. Garnett, eds.) 2622\u20132630. Curran Associates, Red Hook."},{"key":"69","unstructured":"ZHANG, Y.<\/small>, MONDELLI, M.<\/small> and VENKATARAMANAN, R.<\/small> (2022). Precise asymptotics for spectral methods in mixed generalized linear models. arXiv preprint, Available at arXiv:2211.11368<\/a>."},{"key":"3","unstructured":"ARNABOLDI, L.<\/small>, STEPHAN, L.<\/small>, KRZAKALA, F.<\/small> and LOUREIRO, B.<\/small> (2023). From high-dimensional & mean-field dynamics to dimensionless ODEs: A unifying approach to SGD in two-layers networks. In Proceedings of Thirty Sixth Conference on Learning Theory<\/i> (G. Neu and L. Rosasco, eds.). Proceedings of Machine Learning Research<\/i> 195<\/b> 1199\u20131227. PMLR."},{"key":"19","unstructured":"COHEN, J.<\/small>, KAUR, S.<\/small>, LI, Y.<\/small>, KOLTER, J. Z.<\/small> and TALWALKAR, A.<\/small> (2021). Gradient descent on neural networks typically occurs at the edge of stability. In International Conference on Learning Representations<\/i>."},{"key":"21","unstructured":"DAMIAN, A.<\/small>, LEE, J.<\/small> and SOLTANOLKOTABI, M.<\/small> (2022). Neural networks can learn representations with gradient descent. In Proceedings of Thirty Fifth Conference on Learning Theory<\/i> (P.-L. Loh and M. Raginsky, eds.). Proceedings of Machine Learning Research<\/i> 178<\/b> 5413\u20135452. PMLR."},{"key":"28","unstructured":"GHORBANI, B.<\/small>, KRISHNAN, S.<\/small> and XIAO, Y.<\/small> (2019). An investigation into neural net optimization via Hessian eigenvalue density. In Proceedings of the<\/i> 36th International Conference on Machine Learning<\/i> (K. Chaudhuri and R. Salakhutdinov, eds.). Proceedings of Machine Learning Research<\/i> 97<\/b> 2232\u20132241. PMLR."},{"key":"30","unstructured":"GUR-ARI, G.<\/small>, ROBERTS, D. A.<\/small> and DYER, E.<\/small> (2019). Gradient Descent Happens in a Tiny Subspace."},{"key":"32","unstructured":"JACOT, A.<\/small>, GABRIEL, F.<\/small> and HONGLER, C.<\/small> (2020). The asymptotic spectrum of the Hessian of DNN throughout training. In International Conference on Learning Representations<\/i>."},{"key":"38","unstructured":"LIAO, Z.<\/small> and MAHONEY, M. W.<\/small> (2021). Hessian eigenspectra of more realistic nonlinear models. In Advances in Neural Information Processing Systems<\/i> (A. Beygelzimer, Y. Dauphin, P. Liang and J. W. Vaughan, eds.)."},{"key":"41","unstructured":"MAILLARD, A.<\/small>, BEN AROUS, G.<\/small> and BIROLI, G.<\/small> (2020). Landscape complexity for the empirical risk of generalized linear models. In Proceedings of the First Mathematical and Scientific Machine Learning Conference<\/i> (J. Lu and R. Ward, eds.). Proceedings of Machine Learning Research<\/i> 107<\/b> 287\u2013327. PMLR."},{"key":"44","unstructured":"MARTIN, C. H.<\/small> and MAHONEY, M. W.<\/small> (2019). Traditional and Heavy Tailed Self Regularization in Neural Network Models."},{"key":"47","unstructured":"MOUSAVI-HOSSEINI, A.<\/small>, PARK, S.<\/small>, GIROTTI, M.<\/small>, MITLIAGKAS, I.<\/small> and ERDOGDU, M. A.<\/small> (2023). Neural networks efficiently learn low-dimensional representations with SGD. In The Eleventh International Conference on Learning Representations<\/i>."},{"key":"51","unstructured":"PAPYAN, V.<\/small> (2019). Measurements of three-level hierarchical structure in the outliers in the spectrum of deepnet hessians. In Proceedings of the<\/i> 36th International Conference on Machine Learning<\/i> (K. Chaudhuri and R. Salakhutdinov, eds.). Proceedings of Machine Learning Research<\/i> 97<\/b> 5012\u20135021. PMLR."},{"key":"53","doi-asserted-by":"publisher","unstructured":"PAPYAN, V.<\/small>, HAN, X. Y.<\/small> and DONOHO, D. L.<\/small> (2020). Prevalence of neural collapse during the terminal phase of deep learning training. Proc. Natl. Acad. Sci. USA<\/i> 117<\/b> 24652\u201324663.","DOI":"10.1073\/pnas.2015509117"},{"key":"54","unstructured":"PAQUETTE, C.<\/small>, LEE, K.<\/small>, PEDREGOSA, F.<\/small> and PAQUETTE, E.<\/small> (2021). SGD in the large: Average-case analysis, asymptotics, and stepsize criticality. In Conference on Learning Theory<\/i> 3548\u20133626. PMLR."},{"key":"55","unstructured":"PENNINGTON, J.<\/small> and BAHRI, Y.<\/small> (2017). Geometry of neural network loss surfaces via random matrix theory. In Proceedings of the<\/i> 34th International Conference on Machine Learning<\/i> (D. Precup and Y. W. Teh, eds.). Proceedings of Machine Learning Research<\/i> 70<\/b> 2798\u20132806. PMLR."},{"key":"57","unstructured":"REFINETTI, M.<\/small>, GOLDT, S.<\/small>, KRZAKALA, F.<\/small> and ZDEBOROV\u00c1, L.<\/small> (2021). Classifying high-dimensional Gaussian mixtures: Where kernel methods fail and neural networks succeed. In International Conference on Machine Learning<\/i> 8936\u20138947. PMLR."},{"key":"58","unstructured":"SAAD, D.<\/small> and SOLLA, S.<\/small> (1995). Dynamics of on-line gradient descent learning for multilayer neural networks. Adv. Neural Inf. Process. Syst.<\/i> 8<\/b>."},{"key":"59","doi-asserted-by":"crossref","unstructured":"SAAD, D.<\/small> and SOLLA, S. A.<\/small> (1995). On-line learning in soft committee machines. Phys. Rev. E<\/i> 52<\/b> 4225.","DOI":"10.1103\/PhysRevE.52.4225"},{"key":"61","unstructured":"SAGUN, L.<\/small>, EVCI, U.<\/small>, G\u00dcNEY, V. U.<\/small>, DAUPHIN, Y. N.<\/small> and BOTTOU, L.<\/small> (2017). Empirical analysis of the Hessian of over-parametrized neural networks. CoRR. Available at arXiv:1706.04454<\/a>."},{"key":"67","unstructured":"XIE, Z.<\/small>, TANG, Q.-Y.<\/small>, SUN, M.<\/small> and LI, P.<\/small> (2023). On the overlooked structure of stochastic gradients. In Thirty-Seventh Conference on Neural Information Processing Systems<\/i>."}],"container-title":["The Annals of Statistics"],"original-title":[],"deposited":{"date-parts":[[2026,10,6]],"date-time":"2026-10-06T21:12:13Z","timestamp":1791321133000},"score":1,"resource":{"primary":{"URL":"https:\/\/projecteuclid.org\/journals\/annals-of-statistics\/volume-54\/issue-4\/Local-geometry-of-high-dimensional-mixture-models--Effective-spectral\/10.1214\/26-AOS2621.full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,1]]},"references-count":69,"journal-issue":{"issue":"4","published-online":{"date-parts":[[2026,8,1]]}},"URL":"https:\/\/doi.org\/10.1214\/26-aos2621","relation":{},"ISSN":["0090-5364","2168-8966"],"issn-type":[{"value":"0090-5364","type":"print"},{"value":"2168-8966","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,1]]}}}