Abstract:Log-concave sampling has witnessed remarkable algorithmic advances in recent years, but the corresponding problem of proving lower bounds for this task has remained elusive, with lower bounds previously known only in dimension one. In this work, we establish the following query lower bounds: (1) sampling from strongly log-concave and log-smooth distributions in dimension $d\ge 2$ requires $\Omega(\log \kappa)$ queries, which is sharp in any constant dimension, and (2) sampling from Gaussians in dimension $d$ (hence also from general log-concave and log-smooth distributions in dimension $d$) requires $\widetilde \Omega(\min(\sqrt\kappa \log d, d))$ queries, which is nearly sharp for the class of Gaussians. Here $\kappa$ denotes the condition number of the target distribution. Our proofs rely upon (1) a multiscale construction inspired by work on the Kakeya conjecture in geometric measure theory, and (2) a novel reduction that demonstrates that block Krylov algorithms are optimal for this problem, as well as connections to lower bound techniques based on Wishart matrices developed in the matrix-vector query literature.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is related to the lower bound of query complexity for sampling from log - concave distributions. Specifically, the authors focus on the minimum number of queries required for sampling from strongly log - concave and log - smooth distributions in high - dimensional spaces and attempt to prove these lower bounds. ### Problem Background In recent years, significant algorithmic progress has been made in the field of log - concave sampling, but challenges still remain in proving the lower bounds of query complexity for such tasks. Previously known lower bounds were limited to the one - dimensional case. In order to more comprehensively understand the complexity of log - concave sampling, especially its performance in high - dimensional spaces, the authors strive to establish lower bounds of query complexity in higher dimensions. ### Main Contributions 1. **Lower Bounds in Low - Dimensional Spaces**: - The authors prove that in any constant dimension $d \geq 2$, sampling from a $d$-dimensional log - concave distribution with a condition number of $\kappa$ requires at least $\Omega(\log \kappa)$ queries. - This result is stronger than the previous lower bound of $\Omega(\log \log \kappa)$ in the one - dimensional case, and when the dimension is fixed, they also give a matching upper bound $O(\log \kappa)$, thus solving the complexity problem of log - concave sampling in constant dimensions. 2. **Lower Bounds in High - Dimensional Spaces**: - For high - dimensional spaces, the authors prove that sampling from a $d$-dimensional Gaussian distribution with a condition number of $\kappa$ requires at least $\Omega(\min(\sqrt{\kappa}, d))$ queries. - In particular, for sufficiently large $d$ and $\kappa \leq d^{1/5 - \delta}$, they further prove that the number of queries required for sampling is $\Omega_\delta(\sqrt{\kappa} \log d)$, which matches the upper bound of the block Krylov method. ### Technical Means - **Geometric Construction**: The authors introduce a multi - scale geometric construction method inspired by the Kakeya conjecture to generate families of distributions that satisfy specific conditions. - **Wishart Matrix**: By using the properties of the Wishart matrix, the authors show how to reduce the inverse - trace estimation problem to the sampling problem, thereby establishing the lower bound for sampling high - dimensional Gaussian distributions. - **Block Krylov Algorithm**: The authors prove the optimality of the block Krylov algorithm for such problems and further strengthen the lower - bound results through reduction techniques. ### Conclusion By introducing new techniques and construction methods, this paper establishes the lower bounds of query complexity for log - concave sampling in multiple dimensions, fills the gaps in this field, and provides an important theoretical basis for future research.

Query lower bounds for log-concave sampling

Adaptive complexity of log-concave sampling

On theoretical guarantees and a blessing of dimensionality for nonconvex sampling

Quantum Algorithms for Sampling Log-Concave Distributions and Estimating Normalizing Constants.

Log-Concave Sampling on Compact Supports: A Versatile Proximal Framework

Quantum Lower Bounds by Sample-to-Query Lifting

Faster Sampling from Log-Concave Densities over Polytopes via Efficient Linear Solvers

SQ Lower Bounds for Learning Bounded Covariance GMMs

Statistical-Computational Trade-offs for Density Estimation

New Lower Bounds for Testing Monotonicity and Log Concavity of Distributions

Double Randomized Underdamped Langevin with Dimension-Independent Convergence Guarantee.

Non-Stochastic CDF Estimation Using Threshold Queries

Rényi-infinity constrained sampling with $d^3$ membership queries

Statistical Query Lower Bounds for Learning Truncated Gaussians

Complexity of zigzag sampling algorithm for strongly log-concave distributions

Log-concave Sampling from a Convex Body with a Barrier: a Robust and Unified Dikin Walk

Discrete logarithmic Sobolev inequalities in Banach spaces

Better and Simpler Lower Bounds for Differentially Private Statistical Estimation

Zeroth-Order Sampling Methods for Non-Log-Concave Distributions: Alleviating Metastability by Denoising Diffusion

Convergence of Unadjusted Langevin in High Dimensions: Delocalization of Bias

Stochastic Quantum Sampling for Non-Logconcave Distributions and Estimating Partition Functions