Papers for
optimization algorithm developers
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Neural network size estimates improved by counting parameter symmetries
Sharp Covering-Number Bounds for Feedforward ReLU Networks via Parameter Symmetries
Abstract: In neural networks, distinct parameter configurations often realize the same input--output mapping, indicating that nominal parameter counts can overstate functional complexity. In this paper, we characterize the parameter-space symmetries of ReLU and Leaky ReLU networks across fully connected, convolutional, and residual architectures. We prove that every quasianalytic parameter map preserving the network function is a composition of positive channel scaling and channel permutation. By quotienting out both symmetries, we derive uniform covering-number upper bounds in which positive scaling reduces the complexity exponent from the total parameter count $S$ to $S-H$ (with $H$ denoting the number of scalable channels) and permutation symmetry divides the bound by the factorial count $G$ of channel permutations. We establish matching lower bounds showing that the exponent $S-H$ is sharp in general and that the divisor $G$ is attained for one-hidden-layer and generically identifiable networks. We further extend these results to general continuous piecewise-linear activated networks. Finally, we apply our covering-number bounds to derive statistical risk guarantees for empirical risk minimization and capacity control beyond i.i.d.\ settings, and show how the underlying symmetries help explain weight rebalancing in function approximation and flat directions in empirical risk optimization.
Dual-space preconditioning improves solving variational inequalities and roots
Dual-Space Preconditioning for Variational Inequalities and Root-Finding Problems
Abstract: This paper develops a dual-space preconditioning framework for variational inequalities and root-finding problems beyond classical cocoercivity and Lipschitz continuity assumptions. To this end, we introduce both a novel cocoercivity-type condition and a relaxed Lipschitz continuity condition, and derive convergence of deterministic and stochastic methods under these generalized conditions. For isotropic preconditioners, the resulting algorithms can be interpreted as variable-stepsize variants of classical methods, extending recent analyses based on $(L_0, L_1)$-type growth conditions while allowing simpler proofs and larger stepsize ranges. We also show that the framework naturally handles variational inequalities through weak Minty-type arguments. Finally, we study the associated continuous-time dynamics and derive a variational and optimal control interpretation of the preconditioned flow. Overall, the results show that dual-space preconditioning provides a flexible mechanism for designing and analyzing operator methods beyond the standard Euclidean Lipschitz setting.
Finite spectral depth changes early neural net training paths
Finite Spectral Depth Controls Gaussian Entrance Rank and Avoidance of the First Gradient-Flow Saddle
Abstract: Matrix factorization provides a simple model for studying the training dynamics of neural networks. Recent studies of gradient flow under small initialization have analyzed how trajectories escape from a neighborhood of the origin and approach their first nonzero stationary point. In this paper, we study continuous-time dynamics that applies a finite number of Newton-Schulz iterations to the normalized matrix gradient (Muon). We show that the number of Newton-Schulz steps can change the rank selected during the early stages of training and allow the resulting trajectory to bypass the first nonzero stationary point approached by gradient flow. Under an additional residual-inertia condition, we prove that bypass occurs for all sufficiently large finite Newton-Schulz depths and provide an explicit sufficient threshold. We also construct an example in which the bypass threshold is four.
New equations better capture real gradient descent learning dynamics
Stochastic learning flow: moment equations for stochastic gradient descent at finite step size
Abstract: Analytical theories of learning in deep networks often assume gradient flow, discarding the finite step size and batch noise of stochastic gradient descent (SGD). Starting from the discrete update, we derive continuous-time equations for the mean and covariance of the parameter distribution under a Gaussian closure for a general stochastic learning, to first order in the learning rate $η$. Applying these stochastic learning flow (SLF) equations to SGD, we show that the mean descends an effective loss regularized by the covariance and loss curvature, while the covariance, unlike in stochastic differential equation approximations, becomes unstable at curvature $2/η$. In SGD ensembles of linear and nonlinear networks, SLF predicts learning curves and mean trajectories more accurately than gradient flow and captures the covariance dynamics, failing where the parameter distribution becomes strongly non-Gaussian. In the full-batch limit, the edge of stability corresponds to a stationary point of the coupled sharpness-covariance dynamics, recovering self-stabilization and central flow, and batch-to-batch curvature fluctuations lower this threshold under SGD.
Improved newton boosting method with better convergence and adaptivity
Affine-invariant Cubic Newton with Weak Learners
Abstract: Modern gradient boosting algorithms are based on Newton's method, which excels empirically but lacks global convergence guarantees. While recent work on Gradient-Regularized Newton (GRN) boosting achieved a global $\mathcal{O}(1/γ^{12}k^2)$ rate, its analysis was not able to exploit the natural Hessian geometry and suffers from a heavy $γ^{12}$ dependence on a worst-case `weak gradient edge' $γ$. In this paper, we tackle these theoretical drawbacks by generalizing the Affine-Invariant Cubic Newton (AICN) scheme to convex optimization with weak learners. Our analysis is based on average cosine values $\overlineΘ_k$ along the optimization path and directional improvements $δ_k$ in the Hessian-induced norm. We show that the weak-AICN update reduces to a damped weak-Newton step with an explicit stepsize determined by $δ_k$. Under semi-strong self-concordance, we establish a global $\mathcal{O}(1/\overlineΘ_k^4 k^2)$ rate, significantly improving upon the weak-GRN bound. We further prove a local linear rate, global linear convergence for Hessian-dominated losses, cosine angle lower bounds, and present weak-AdaAICN, which adaptively tunes the cubic regularization. Finally, we demonstrate the broader applicability of this framework beyond gradient boosting by showing that Newton coordinate descent emerges naturally as a special case of optimization with weak learners.
Stochastic gradient descent converges in deep linear neural networks
Convergence of stochastic gradient descent for learning linear neural networks
Abstract: Stochastic gradient descent (SGD) is the cornerstone of modern deep learning, yet its theoretical convergence properties remain poorly understood in non-convex settings. To address this gap, we study SGD dynamics in deep linear neural networks - simplified models that retain key non-convex characteristics while remaining analytically tractable. Building on recent results for gradient descent, we extend convergence guarantees to SGD under regularized squared loss. We establish almost sure boundedness of SGD iterates. Using the framework of asymptotic pseudotrajectories from stochastic approximation theory, our analysis establishes then, for both deterministic and adaptive step-size choices, convergence of SGD to the set of critical points under mild conditions (without relying on Lipschitz properties or whitened data). To our knowledge, this is the first work providing almost sure convergence guarantees for SGD in deep linear networks, bridging the gap between theoretical optimization and practical stochastic training.
Deterministic optimization costs scale poorly under low curvature trace
Lower Bounds for Smooth Nonconvex Zeroth-Order Optimization: The Price of Determinism under Low Curvature Trace
Abstract: Low curvature trace, an effective rank condition, is a common explanation for why randomized zeroth-order methods scale to high-dimensional tasks such as language model fine-tuning. Can deterministic methods achieve the same savings? We show that they cannot. We study finding $\varepsilon$-stationary points of $L$-smooth nonconvex functions on $\mathbb{R}^d$ from exact function values, assuming that below the starting value the Hessian is dominated by a fixed matrix with trace at most $LE$; this is the effective rank condition used to explain zeroth-order fine-tuning. When $ET\le d$, the Gaussian two-point method that such fine-tuning uses needs only $O(ET)$ calls if it rejects steps that raise $f$, where $T=LΔ/\varepsilon^2$ is the iteration count of gradient descent from initial gap $Δ$, and we prove that every randomized method needs $Ω(ET)$ calls, so this cost is optimal up to constants. In contrast, every deterministic method needs $\widetildeΩ(d\min\{T,\sqrt{ET}\})$ calls, and adaptive subspace search attains $O(d\min\{T,\sqrt{ET}\})$: a deterministic method pays about $d$ calls per discovered direction, and the trace budget allows about $\sqrt{ET}$ directions. Determinism thus multiplies the cost by a factor of order $d/(E+\sqrt{ET})$ up to logarithmic factors, which grows linearly with $d$ for fixed $E$ and $T$; once $E$ exceeds $T$ by a logarithmic factor, deterministic methods need $Ω(dT)$ calls, as without any trace budget. Our lower bounds hold for $C^\infty$ functions with the matrix bound on all of $\mathbb{R}^d$, even when each call also returns gradient projections priced by rank.
Two hidden layers suffice to compute maximum of many inputs
Two hidden layer suffice: the caterpillar net
Abstract: Every continuous piecewise linear (CPWL) function is a finite signed sum of maxima of affine functions, making maximum representation a central problem in the study of the expressiveness of ReLU networks. A now-disproved conjecture predicted logarithmic depth for maxima. Inspired by the proof by Brunck et al. (2026) that two hidden layers suffice for general CPWL functions, we give the Caterpillar Net, a nonrecursive construction for the maximum of $N$ inputs with two hidden layers of widths $O(N^2)$ and $2^{O(N\log N)}$. The construction has a direct geometric interpretation.
Deep network training converges better with noise and wider layers
Global Convergence in Deep Networks via the Second Law of Thermodynamics
Abstract: Existing convergence analyses of finite-width overparameterized deep neural networks under standard parameterization often rely on parameters remaining close to their initialization and typically neglect the role of noise, despite empirical evidence that stochasticity can improve optimization and generalization. To bridge this gap, we study $L_2$-regularized continuous-time Langevin dynamics (CLD) with isotropic noise, for training feedforward networks with smooth activations. We introduce a change-of-measure inequality, inspired by the second law of thermodynamics, that enables control of key properties of the training dynamics under the time-evolving parameter distribution. Leveraging a sharpened lower bound on the minimum eigenvalue of the Neural Tangent Kernel (NTK) at initialization, we show that injected noise helps preserve NTK stability with high probability without confining the weights to a small neighborhood of their initialization. As a result, we derive a non-asymptotic upper bound on the expected training loss. Our analysis yields explicit conditions on width and noise temperature that guarantee convergence to a finite loss floor. Finally, we experimentally show the network Jacobian can change from initialization, departing the commonly analyzed lazy regime.
Zeroth-order nonconvex optimization lower bounds match theory limits
On Global Joint Lower Bounds in Zeroth-Order Nonconvex Optimization
Abstract: Recent work has sought to transfer first-order nonconvex optimization guarantees to zeroth-order methods through gradient estimators, typically incurring a factor linear in $d$. \citet{kornowski2024} obtained an $\mathcal{O}(dδ^{-1}ε^{-3})$ oracle complexity for nonsmooth nonconvex optimization and asked whether its dependence on all three parameters is jointly optimal. Although the dimension dependence and the first-order dependence on $(δ,ε)$ are separately known to be unavoidable, a matching zeroth-order lower bound has remained open. We answer this question in the unrestricted full-space setting. For smooth stochastic nonconvex optimization, we establish a scalar-oracle lower bound of $Ω\!\left(dL_0^2\min\{ΔH,L_0^2\}/ε^4\right)$ assuming a smooth, globally lower-bounded population objective and a second-moment bound on sample Lipschitz constants. A smooth reduction then yields a lower bound for Goldstein stationarity, proving the joint minimax optimality of the dependence on $(d,δ,ε)$ under explicit high-dimensional and small-accuracy conditions. Crucially, our results allow algorithms to query arbitrary points in $\mathbb{R}^d$, adaptively select sample-function labels, reuse the same sampled function at any number of locations, and output points without spatial restrictions. Thus, unlike lower bounds based on bounded query regions or fresh evaluation noise, our guarantees apply to a fully adaptive, reusable-sample scalar oracle. We also establish a separate noiseless exact-value lower bound against randomized algorithms, and hence also against deterministic methods. Our proof combines a globally defined sequential nonconvex construction with an exact Gaussian scalar channel, multistage random-field estimates, and an antithetic finite-sample realization that preserves information hardness under adaptive reuse.
Randomized methods have provable limits in convex set optimization
Lower Bounds for Randomized Linear-Oracle Optimization over Strongly Convex Sets
Abstract: Recent results [arXiv:2602.04378, arXiv:2602.22608] have established worst-case objective-error lower bounds of order $1/T^2$ after $T$ linear minimization oracle (LMO) calls, matching the upper bound [arXiv:1406.1305] in its dependence on $T$ and thereby resolving a long-standing open question. Upon close inspection, however, both lower bounds are somewhat brittle. The first [arXiv:2602.04378] applies only to the standard Frank--Wolfe algorithm, and, as we show, its prescribed slow exact-line-search trajectories can amplify sufficiently small initialization errors exponentially, corresponding to roughly one additional bit of initialization accuracy per iteration. The second [arXiv:2602.22608] obtains lower bounds for any deterministic LMO-based first-order method via a resisting oracle that adversarially adapts to the queries of the algorithm. In both cases it is conceivable that randomization may circumvent these lower bounds; in particular, neither establishes a lower bound for unrestricted randomized first-order LMO methods. We close this gap by proving an $Ω(L\operatorname{diam}(S)^2/T^2)$ distributional lower bound for randomized first-order linear minimization oracle methods over strongly convex sets. While our construction reuses the family of [arXiv:2602.22608], our proof characterizes the conditional distribution of the hidden permutation after adaptive queries and bounds the expected decrease of a combinatorial statistic controlling optimizer variance. We further show that the same error threshold holds with probability at least $3/4$ and yields a stopping-time tail bound.
Comb inequality violation detection proven computationally hard
Separating comb inequalities is NP-hard
Abstract: Comb inequalities are important cutting planes for the symmetric travelling salesman problem, yet the complexity of their exact separation has remained a longstanding question in polyhedral combinatorics. We present a reduction from 3-SAT proving that deciding whether a comb inequality is violated is NP-complete and that the corresponding separation problem is NP-hard. This result holds even when the input vector belongs to the subtour elimination polytope, every edge value is zero, one half or one and the support graph is nonplanar with maximum degree four. The reduction constructs a graph in which six-vertex ladder gadgets encode Boolean relations and cubic graphs enforce consistency among occurrences of each logical variable. For a propositional logic formula with $v$ variables and $m$ clauses, the constructed graph has $40v+70m+30$ vertices and a linear number of positive edges. The reduction also proves hardness when every permitted tooth has two or four vertices with half of the tooth in the handle. We discuss consequences for approximating the maximum comb violation, optimization over the comb relaxation and explain why hardness of separation does not automatically transfer to larger inequality families.
Undercomplete linear autoencoders favor large decoder weights in training
Broken scale symmetries in undercomplete linear autoencoders
Abstract: Neural network loss landscapes have many symmetries, which are preserved by gradient flow but broken by finite-stepsize stochastic gradient descent (SGD). A canonical example of such a symmetry is scale in homogeneous networks: one can scale up the parameters in one layer and down in the next without changing the network output. Previous work has documented cases in which SGD breaks this symmetry in favor of balancing gradient noise or minimizing fluctuations. Here, we show that the solution geometry of undercomplete linear autoencoders instead selects a preferred sign for scale drift: on the PCA solution manifold, SGD favors large decoder weights. This directed scale drift occurs on a slow timescale, and its dynamics admit an analytically-tractable effective description. However, it cannot continue indefinitely: increasing scale eventually drives the dynamics towards a finite-stepsize stability boundary. The resulting solutions are sharper than a balanced baseline in the sense of the maximum eigenvalue of the loss Hessian, but different sharpness measures can move in opposing directions. Thus, undercomplete autoencoders give a concrete illustration of how loss geometry can convert residual gradient noise into directed motion along a manifold of functionally-equivalent solutions.
Decoded quantum interferometry breaks topological barriers in sampling
Gibbs Sampling in the Shattered Phase by Decoded Quantum Interferometry
Abstract: We apply Decoded Quantum Interferometry (DQI) to sample from the Gibbs measures of classical Ising spin Hamiltonians. We show that this Gibbs sampling problem reduces to a quantum decoding problem, and the temperature achievable by DQI is determined by the performance of decoding algorithms. We then focus on the task of Gibbs sampling for classical Ising $k$-spin glasses (or Max-$k$-XORSAT) on random Erdős-Rényi hypergraphs with average degree $D\ge k$. In a temperature range beginning asymptotically at the predicted dynamical phase transition, $β_{\rm dyn}(k,D) = \sqrt{(2\ln k)/D}\times [1+o_{k\to\infty}(1)]$, we show that shattering and disorder chaos form a topological barrier that obstructs many algorithms, including Glauber dynamics and any algorithm whose output distribution is "stable" under perturbations of the input. In contrast, we prove that this barrier can be broken both by a classical algorithm based on Prange's method, and by DQI equipped with a quantum decoder. For example, when $D=αk$ with fixed $α>1$, both Prange's algorithm and DQI can sample at any inverse temperature $β< \tanh^{-1}(1/α)$ for sufficiently large $k$, well beyond the dynamical threshold $β_{\rm dyn} \sim \sqrt{2\ln k / (αk)}$. Therefore, our results show that DQI can overcome topological barriers that obstruct stable algorithms.
Weak proximal oracles trade speed for complexity in convex optimization
Complexities of Weak Proximal Oracle Methods for Composite Convex Optimization
Abstract: We consider a standard convex composite optimization problem with either smooth or nonsmooth objective function, and under quadratic growth. In recent years, several works gave algorithms based on a \textit{weak proximal oracle} (WPO) that essentially match in oracle complexities proximal (sub)gradient methods relying on exact prox operations. Importantly, such WPOs, which relax the strong optimality condition of the standard prox operator, may admit much more efficient implementation in terms of runtime when optimal solutions have some sparse structure. A question remained if such WPO-based methods can be accelerated (in the sense of Nesterov's accelerated gradient). In this work we provide a negative answer by establishing lower bounds against both deterministic and randomized methods. Thus, while WPOs can substantially reduce the cost of individual oracle calls, this comes with an inherent loss in oracle complexity. We also provide a new upper-bound for WPO-based nonsmooth convex composite optimization, nearly matching the proximal subgradient method.
Connections between quadratic transform and schur complement revealed
Connections Between Quadratic Transform for Fractional Programming and Schur Complement
Abstract: This paper shows that there are intimate connections between the quadratic transform technique for solving fractional programming (FP) problems and the Schur-complement technique in matrix analysis. We demonstrate that the quadratic transform technique is related to two aspects of the Schur complement: (i) the linear matrix inequality (LMI) condition for positive semidefiniteness and (ii) the matrix determinant formula. Specifically, we establish that the quadratic transform and the Schur-complement LMI condition imply each other. This connection allows us to provide new interpretations of the auxiliary variable in the quadratic transform, and it allows us to rederive the Schur-complement determinant formula. Furthermore, this connection leads to generalizations of the quadratic transform in FP and the Schur-complement LMI that can accommodate generalized matrix inverse. As an application in information theory, we apply the generalized FP framework to the least-favorable-noise minimax formulation of the Gaussian vector broadcast channel sum capacity problem. When the least-favorable noise covariance is singular, matrix-inverse-based Karush-Kuhn-Tucker (KKT) analysis would require a careful analysis of the input and output spaces of the channel. We show using generalized FP that an auxiliary-variable representation of the singular matrix fraction directly yields the reciprocal multiple-access channel and recovers the uplink-downlink duality relation for sum capacity.