arXiv subjects

Brian Bullins

Publications and source records attributed to Brian Bullins.

At least 19 recordsLinked to original sources

arXiv2026

MuonIO: Principled Norm-Aware Descent for Embedding Tables and Language Model Heads

The Muon optimizer derives its update rule for hidden linear layers by solving a local linearization of the loss penalized by the spectral norm, motivated by an RMS-stability argument for dense linear layers. Standard Muon implementations, however, exclude the input (embedding table) and output (language model head) layers from this principled treatment, for which they use AdamW instead. We present MuonIO, a single Muon-style update for both of these layers. For the language model head $\mathbf{L} \in \mathbb{R}^{V \times d}$, we motivate the use of the $2\to\infty$ operator norm, due to the Lipschitz continuity of the softmax output geometry, while for the embedding table $\mathbf{E} \in \mathbb{R}^{d \times V}$, we draw on the $1 \to 2$ operator norm, based on the one-hot input geometry identified by Bernstein & Newhouse (2025). The identity $\lVert\mathbf{L}\rVert_{2\to\infty}=\lVert\mathbf{L}^\top\rVert_{1\to2}$ then puts both matrices in the same vocabulary-oriented geometry: MuonIO applies a single normalized-vector rule, which appears as column normalization for $\mathbf{E}$ and row normalization for $\mathbf{L}$. Empirical evaluations demonstrate the effectiveness of our approach, with MuonIO reducing I/O optimizer state memory by 50% and I/O update FLOPs by $\sim$46% compared to Muon for 1B LLaMA pretraining on C4, while also improving validation perplexity.

cs.LG↗

arXiv2026

Spectral Saliency for Machine Unlearning

Machine unlearning (MU) aims to remove the influence of specific training data while preserving model utility. As the name suggests, MU can be viewed as the inverse of learning, using gradient-based updates to reduce the influence of a forget-set by counteracting the previously learned behavior. Recently, Muon, a gradient descent variant, has been introduced. Muon applies spectral magnitude normalization to encourage exploration of rare directions and demonstrates promising performance. Inspired by Muon, we adopt the spectral view for unlearning and propose Spectral Saliency Unlearning (SSU). SSU thresholds weak singular components and updates only those directions supported by a confident unlearning signal. We further provide theoretical justification for this thresholding approach from the perspective of the forgetting-retention trade-off. Experiments across image classifiers, diffusion models, and LLMs demonstrate SSU's effectiveness.

cs.LG↗

arXiv2026

Near-Optimal Acceleration for Smooth $\ell_p$ / $\ell_q$ Nondual Convex First-Order Oracle Optimization

We study the optimization of convex objectives with $(L,κ-1)$-Hölder-continuous gradients in $\ell_q$ over $R B_p^d$, $1<κ\le 2$. (MG26) provides selectors with a movement bound for the problem of chasing high-dimensional convex nested sets for every $p<q$ and generally reduces Lipschitz convex optimization to bounds on the movement of selectors. We couple that movement with Hölder descent yielding a polynomial-runtime first-order method whose feasible output, in the high-dimensional regime $T\le d$ and for $p<\min\{q,2\}$, has error $$ \widetilde O_{κ,p,q}\!\left( \frac{LR^κ}{T^{κ(1+1/p-(1/q-1/2)_+)-1}} \right), $$ after $T$ queries to a first-order oracle, solving the COLT 2015 open problem of (Guz15), up to logarithmic factors. At $(p,q)=(1,2)$, the rate is $\widetilde{O}(LR^κ/T^{2κ-1})$, including $\widetilde{O}(LR^2/T^{3})$ cubic decay in the smooth case.

math.OC↗

arXiv2026

The First-Order Oracle Complexity of Lipschitz Convex Optimization in Nondual Settings

We study first-order black-box convex optimization over an $\ell_p$-ball for objectives Lipschitz in the $\ell_q$-norm, solving in the affirmative the nonsmooth version of the COLT open question (Guz15b) on whether the geometry of a smaller feasible set ($p < q$) can improve convergence rates in convex optimization, and matching prior lower bounds up to logarithmic factors. Our rates include $\widetilde O(1/T)$ for convex Euclidean-Lipschitz optimization over the $\ell_1$-ball, improving on the $O(1/\sqrt{T})$ classical rate under general assumptions. The key technical device is a new online learning game, where the comparator is evaluated using the maximum of affine losses observed so far. We bound the value of this game above and below in terms of a combinatorial online learning quantity: the sequential fat-shattering dimension, which we characterize for the $\ell_p / \ell_q$ case. Our results generally apply when the feasible set $X$ and the set of possible subgradients $H$ are convex, centrally symmetric, and admit a type of minmax theorem, advancing on a fundamental question by Sridharan [Sri12, Section 10.1.2, Q3]. As a geometric consequence of our analysis, of independent interest, we obtain estimates for the expected distance of a convex hull of samples to their mean in several Banach geometries, a version of the celebrated Wendel's theorem (Wen62), but quantitative and for bounded general distributions as opposed to centrally symmetric ones.

math.OC↗

arXiv2026

On the Limits of Biased Derivative Information for Nonconvex Stochastic Optimization

We consider the problem of finding $δ$-stationary points for $δ> 0$, i.e., $x \in \mathbb{R}^d$ such that $||\nabla F(x)|| \le δ$, for smooth, non-convex objectives, where the derivative oracles are not only stochastic but also biased. In the first-order setting, we provide tight lower bounds for finding an $O((ε+ B^2)^{1/2})$-stationary point, for $ε> 0$ and where $B$ is a bound on the gradient bias, matching the upper bounds of Ajalloeian and Stich (2020). We then establish bias-dependent lower bounds for algorithms that use higher-order derivative information for finding $O(ε+ B_{\max})$-stationary points, where $B_{\max}$ is a bound on the maximum bias for all derivatives. To complement these lower bounds, we develop trust-region based methods that, for certain ranges of bias, provide guarantees that match the corresponding lower bounds. We further improve upon the oracle complexity in high bias settings through a higher-order variance reduction scheme, in particular demonstrating the benefits, in some cases, of using higher-order derivative information, whereas such improvements are known to be unattainable for stochastic unbiased settings.

math.OC↗

arXiv2026

Decoupling Variance and Scale-Invariant Updates in Adaptive Gradient Descent for Unified Vector and Matrix Optimization

Adaptive methods like Adam have become the $\textit{de facto}$ standard for large-scale vector and Euclidean optimization due to their coordinate-wise adaptation with a second-order nature. More recently, matrix-based spectral optimizers like Muon (Jordan et al., 2024b) show the power of treating weight matrices as matrices rather than long vectors. Linking these is hard because many natural generalizations are not feasible to implement, and we also cannot simply move the Adam adaptation to the matrix spectrum. To address this, we reformulate the AdaGrad update and decompose it into a variance adaptation term and a scale-invariant term. This decoupling produces $\textbf{DeVA}$ ($\textbf{De}$coupled $\textbf{V}$ariance $\textbf{A}$daptation), a framework that bridges between vector-based variance adaptation and matrix spectral optimization, enabling a seamless transition from Adam to adaptive spectral descent. Extensive experiments across language modeling and image classification demonstrate that DeVA consistently outperforms state-of-the-art methods such as Muon and SOAP (Vyas et al., 2024), reducing token usage by around 6.6\%. Theoretically, we show that the variance adaptation term effectively improves the blockwise smoothness, facilitating faster convergence. Our implementation is available at https://github.com/Tsedao/Decoupled-Variance-Adaptation

cs.LG↗

arXiv2026

Can Entry-Wise Clipping Give Spectral Control of Stochastic Gradients?

Training instabilities such as loss spikes are frequently the result of stochastic gradient noise. Because of rare expressions in language training data, and multiple layer composition, the noise impact is heavy-tailed and survives mini-batch averaging. Existing remedies trade off structure against cost: vector-norm clipping ignores the matrix structure of weight updates, while spectral normalization (e.g., Muon (Jordan et al., 2024)) respects it at additional cost. We show that this trade-off can be balanced. Real gradient noise appears to be similar to entry-wise heavy-tailed contamination, and a first-order perturbation analysis reveals a localization property of such noise, under which a simple entry-wise method achieves spectral control. Exploiting this, we derive a tractable surrogate for the Bayes-optimal entry-wise estimator under a Gaussian signal prior. We establish $O(ε^{-4})$ convergence guarantee under Cauchy-contaminated noise. Empirically, we find that smooth shrinkage improves Adam on NanoGPT pretraining, saving ${\sim}7\%$ of training tokens. We further find that applying the entry-wise clipping before spectral normalization yields a ${\sim}2\%$ token saving on top of Muon.

cs.LG↗

arXiv2026

Mirror-Free Proximal Methods

We present a \emph{mirror-free} mirror prox (MFMP) algorithm, which extends the classic approach of Nemirovski (2004) to allow for proximal-like updates without the explicit need for a mirror map. We further analyze the convergence of our method under suitable notions of relative smoothness and relative Lipschitzness, for which we introduce a relaxation of the standard Bregman divergence in terms of more general potential operators. Finally, we show how a strongly monotone variant of our method allows us to solve regularized Taylor-expansion subproblems that appear in both second- and third-order smooth min-max optimization.

math.OC↗

arXiv2026

Beyond First-Order Methods for $\ell_p$-Structured Non-Monotone Variational Inequalities

We propose novel high-order algorithms for a class of $\ell_p$-structured non-monotone variational inequalities. In particular, work by Diakonikolas et al. (2021), which introduced the weak Minty variational inequality (weak-MVI) setting, showed how to find an approximate first-order Euclidean stationary point for a strictly positive range of the weak-MVI parameter $ρ$. However, for the $\ell_p$-norm stationary point setting ($p \neq 2$), their guarantees are limited to $ρ=0$, which recovers the standard MVI setting. In this work, we address this gap by presenting a suite of high-order methods that converge to $\ell_p$-norm stationary points for a suitable range of $ρ> 0$, thereby circumventing previous fundamental challenges in $\ell_p$ settings. We further show convergence for high-order smooth \textit{monotone} operators, generalizing Adil et al. (2022) to the case where $p \geq 2$, and we extend our Euclidean techniques to continuous-time settings.

math.OC↗

arXiv2026

Online Min-Max Optimization: From Individual Regrets to Cumulative Saddle Points

We propose and study an online version of min-max optimization based on cumulative saddle points under a variety of performance measures beyond convex-concave settings. After first observing the incompatibility of (static) Nash equilibrium (SNE-Reg$_T$) with individual regrets even for strongly convex-strongly concave functions, we propose an alternate \emph{static} duality gap (SDual-Gap$_T$) inspired by the online convex optimization (OCO) framework. We provide algorithms that, using a reduction to classic OCO problems, achieve bounds for SDual-Gap$_T$~and a novel \emph{dynamic} saddle point regret (DSP-Reg$_T$), which we suggest naturally represents a min-max version of the dynamic regret in OCO. We derive our bounds for SDual-Gap$_T$~and DSP-Reg$_T$~under strong convexity-strong concavity and a min-max notion of exponential concavity (min-max EC), and in addition we establish a class of functions satisfying min-max EC~that captures a two-player variant of the classic portfolio selection problem. Finally, for a dynamic notion of regret compatible with individual regrets, we derive bounds under a two-sided Polyak-Łojasiewicz (PL) condition.

cs.LG↗

arXiv2026

Convex optimization with $p$-norm oracles

In recent years, there have been significant advances in efficiently solving $\ell_s$-regression using linear system solvers and $\ell_2$-regression [Adil-Kyng-Peng-Sachdeva, J. ACM'24]. Would efficient smoothed $\ell_p$-norm solvers lead to even faster rates for solving $\ell_s$-regression when $2 \leq p < s$? In this paper, we give an affirmative answer to this question and show how to solve $\ell_s$-regression using $\tilde{O}(n^{\fracν{1+ν}})$ iterations of solving smoothed $\ell_p$ regression problems, where $ν:= \frac{1}{p} - \frac{1}{s}$. To obtain this result, we provide improved accelerated rates for convex optimization problems when given access to an $\ell_p^s(λ)$-proximal oracle, which, for a point $c$, returns the solution of the regularized problem $\min_{x} f(x) + λ||x-c||_p^s$. Additionally, we show that these rates for the $\ell_p^s(λ)$-proximal oracle are optimal for algorithms that query in the span of the outputs of the oracle, and we further apply our techniques to settings of high-order and quasi-self-concordant optimization.

math.OC↗

arXiv2025

Balancing Gradient and Hessian Queries in Non-Convex Optimization

We develop optimization methods which offer new trade-offs between the number of gradient and Hessian computations needed to compute the critical point of a non-convex function. We provide a method that for any twice-differentiable $f\colon \mathbb R^d \rightarrow \mathbb R$ with $L_2$-Lipschitz Hessian, input initial point with $Δ$-bounded sub-optimality, and sufficiently small $ε> 0$, outputs an $ε$-critical point, i.e., a point $x$ such that $\|\nabla f(x)\| \leq ε$, using $\tilde{O}(L_2^{1/4} n_H^{-1/2}Δε^{-9/4})$ queries to a gradient oracle and $n_H$ queries to a Hessian oracle for any positive integer $n_H$. As a consequence, we obtain an improved gradient query complexity of $\tilde{O}(d^{1/3}L_2^{1/2}Δε^{-3/2})$ in the case of bounded dimension and of $\tilde{O}(L_2^{3/4}Δ^{3/2}ε^{-9/4})$ in the case where we are allowed only a \emph{single} Hessian query. We obtain these results through a more general algorithm which can handle approximate Hessian computations and recovers the state-of-the-art bound of computing an $ε$-critical point with $O(L_1^{1/2}L_2^{1/4}Δε^{-7/4})$ gradient queries provided that $f$ also has an $L_1$-Lipschitz gradient.

math.OC↗

arXiv2025

Faster Acceleration for Steepest Descent

Recent advances (Sherman, 2017; Sidford and Tian, 2018; Cohen et al., 2021) have overcome the fundamental barrier of dimension dependence in the iteration complexity of solving $\ell_\infty$ regression with first-order methods. Yet it remains unclear to what extent such acceleration can be achieved for general $\ell_p$ smooth functions. In this paper, we propose a new accelerated first-order method for convex optimization under non-Euclidean smoothness assumptions. In contrast to standard acceleration techniques, our approach uses primal-dual iterate sequences taken with respect to $\textit{differing}$ norms, which are then coupled using an $\textit{implicitly}$ determined interpolation parameter. For $\ell_p$ norm smooth problems in $d$ dimensions, our method provides an iteration complexity improvement of up to $O(d^{1-\frac{2}{p}})$ in terms of calls to a first-order oracle, thereby allowing us to circumvent long-standing barriers in accelerated non-Euclidean steepest descent.

math.OC↗

arXiv2025

Tight Lower Bounds under Asymmetric High-Order Hölder Smoothness and Uniform Convexity

In this paper, we provide tight lower bounds for the oracle complexity of minimizing high-order Hölder smooth and uniformly convex functions. Specifically, for a function whose $p^{th}$-order derivatives are Hölder continuous with degree $ν$ and parameter $H$, and that is uniformly convex with degree $q$ and parameter $σ$, we focus on two asymmetric cases: (1) $q > p + ν$, and (2) $q < p+ν$. Given up to $p^{th}$-order oracle access, we establish worst-case oracle complexities of $Ω\left( \left( \frac{H}σ\right)^\frac{2}{3(p+ν)-2}\left( \fracσε\right)^\frac{2(q-p-ν)}{q(3(p+ν)-2)}\right)$ in the first case with an $\ell_\infty$-ball-truncated-Gaussian smoothed hard function and $Ω\left(\left(\frac{H}σ\right)^\frac{2}{3(p+ν)-2}+ \log\log\left(\left(\frac{σ^{p+ν}}{H^q}\right)^\frac{1}{p+ν-q}\frac{1}ε\right)\right)$ in the second case, for reaching an $ε$-approximate solution in terms of the optimality gap. Our analysis generalizes previous lower bounds for functions under first- and second-order smoothness as well as those for uniformly convex functions, and furthermore our results match the corresponding upper bounds in this general setting.

math.OC↗

arXiv2025

Stacey: Promoting Stochastic Steepest Descent via Accelerated $\ell_p$-Smooth Nonconvex Optimization

While popular optimization methods such as SGD, AdamW, and Lion depend on steepest descent updates in either $\ell_2$ or $\ell_\infty$ norms, there remains a critical gap in handling the non-Euclidean structure observed in modern deep networks training. In this work, we address this need by introducing a new accelerated $\ell_p$ steepest descent algorithm, called Stacey, which uses interpolated primal-dual iterate sequences to effectively navigate non-Euclidean smooth optimization tasks. In addition to providing novel theoretical guarantees for the foundations of our algorithm, we empirically compare our approach against these popular methods on tasks including image classification and language model (LLM) pretraining, demonstrating both faster convergence and higher final accuracy. We further evaluate different values of $p$ across various models and datasets, underscoring the importance and efficiency of non-Euclidean approaches over standard Euclidean methods. Code can be found at https://github.com/xinyuluo8561/Stacey .

cs.LG↗

arXiv2025

Model Immunization from a Condition Number Perspective

Model immunization aims to pre-train models that are difficult to fine-tune on harmful tasks while retaining their utility on other non-harmful tasks. Though prior work has shown empirical evidence for immunizing text-to-image models, the key understanding of when immunization is possible and a precise definition of an immunized model remain unclear. In this work, we propose a framework, based on the condition number of a Hessian matrix, to analyze model immunization for linear models. Building on this framework, we design an algorithm with regularization terms to control the resulting condition numbers after pre-training. Empirical results on linear models and non-linear deep-nets demonstrate the effectiveness of the proposed algorithm on model immunization. The code is available at https://github.com/amberyzheng/model-immunization-cond-num.

cs.LG↗

arXiv2023

Federated Composite Saddle Point Optimization

Federated learning (FL) approaches for saddle point problems (SPP) have recently gained in popularity due to the critical role they play in machine learning (ML). Existing works mostly target smooth unconstrained objectives in Euclidean space, whereas ML problems often involve constraints or non-smooth regularization, which results in a need for composite optimization. Addressing these issues, we propose Federated Dual Extrapolation (FeDualEx), an extra-step primal-dual algorithm, which is the first of its kind that encompasses both saddle point optimization and composite objectives under the FL paradigm. Both the convergence analysis and the empirical evaluation demonstrate the effectiveness of FeDualEx in these challenging settings. In addition, even for the sequential version of FeDualEx, we provide rates for the stochastic composite saddle point setting which, to our knowledge, are not found in prior literature.

cs.LG↗

arXiv2023

Beyond first-order methods for non-convex non-concave min-max optimization

We propose a study of structured non-convex non-concave min-max problems which goes beyond standard first-order approaches. Inspired by the tight understanding established in recent works [Adil et al., 2022, Lin and Jordan, 2022b], we develop a suite of higher-order methods which show the improvements attainable beyond the monotone and Minty condition settings. Specifically, we provide a new understanding of the use of discrete-time $p^{th}$-order methods for operator norm minimization in the min-max setting, establishing an $O(1/ε^\frac{2}{p})$ rate to achieve $ε$-approximate stationarity, under the weakened Minty variational inequality condition of Diakonikolas et al. [2021]. We further present a continuous-time analysis alongside rates which match those for the discrete-time setting, and our empirical results highlight the practical benefits of our approach over first-order methods.

math.OC↗