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

📅 2024-09-16
🏛️ International Conference on Learning Representations
📈 Citations: 0
Influential: 0
📄 PDF
🤖 AI Summary
This work establishes tight oracle complexity lower bounds for unconstrained optimization under asymmetric smoothness and convexity assumptions: functions whose $p$-th derivative is $ u$-Hölder continuous (with parameter $H$) and which are $q$-uniformly convex (with parameter $sigma$), where $q eq p + u$. For both regimes—$q > p + u$ and $q < p + u$—we construct novel hard instances via $ell_infty$-ball-truncated Gaussian smoothing. Leveraging high-order Taylor expansions, precise characterizations of uniform convexity, and information-theoretic arguments, we derive tight lower bounds of order $Omegaig(1/varepsilon^{(p+ u)/(q-p)}ig)$. These results unify and generalize existing lower bounds for first- and higher-order smooth/convex optimization, and match the best-known upper bounds exactly. Consequently, the optimal convergence rate for this class of asymmetric optimization problems is fully characterized.

Technology Category

Application Category

📝 Abstract
In this paper, we provide tight lower bounds for the oracle complexity of minimizing high-order H""older smooth and uniformly convex functions. Specifically, for a function whose $p^{th}$-order derivatives are H""older continuous with degree $ u$ and parameter $H$, and that is uniformly convex with degree $q$ and parameter $sigma$, we focus on two asymmetric cases: (1) $q>p + u$, and (2) $q<p+ u$. Given up to $p^{th}$-order oracle access, we establish worst-case oracle complexities of $Omegaleft( left( frac{H}{sigma} ight)^frac{2}{3(p+ u)-2}left( frac{sigma}{epsilon} ight)^frac{2(q-p- u)}{q(3(p+ u)-2)} ight)$ in the first case with an $ell_infty$-ball-truncated-Gaussian smoothed hard function and $Omegaleft(left(frac{H}{sigma} ight)^frac{2}{3(p+ u)-2}+ log^2left(frac{sigma^{p+ u}}{H^q} ight)^frac{1}{p+ u-q} ight)$ in the second case, for reaching an $epsilon$-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 the general setting.
Problem

Research questions and friction points this paper is trying to address.

Tight lower bounds for minimizing high-order Hölder smooth functions
Oracle complexity for uniformly convex functions with asymmetric cases
Generalization of previous bounds for first- and second-order smoothness
Innovation

Methods, ideas, or system contributions that make the work stand out.

Tight lower bounds for high-order Hölder smooth functions
Asymmetric cases analysis with uniform convexity
Generalized lower bounds matching upper bounds
S
Site Bai
Department of Computer Science, Purdue University, West Lafayette, IN, USA
Brian Bullins
Brian Bullins
Assistant Professor, Purdue University
OptimizationMachine Learning