On Learning Parallel Pancakes with Mostly Uniform Weights

Authors: Ilias Diakonikolas, Daniel Kane, Sushrut Karmalkar, Jasper C.H. Lee, Thanasis Pittas

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Theoretical Our first main result is a Statistical Query (SQ) lower bound showing that this quasi-polynomial upper bound is essentially best possible, even for the special case of uniform weights. Specifically, we show that it is SQ-hard to distinguish between such a mixture and the standard Gaussian. We further explore how the distribution of weights affects the complexity of this task. Our second main result is a quasi-polynomial upper bound for the aforementioned testing task when most of the weights are uniform while a small fraction of the weights are potentially arbitrary. This work is theoretical in nature and focuses on advancing fundamental knowledge.
Researcher Affiliation Collaboration 1Department of Computer Sciences, University of Wisconsin Madison, Madison, United States 2University of California San Diego, San Diego, United States 3Microsoft Research, Cambridge, England 4Some of this work was done while the author was a postdoctoral researcher at UW-Madison. 5University of California Davis, Davis, United States. Correspondence to: Thanasis Pittas <EMAIL>.
Pseudocode Yes Algorithm 1 Testing Algorithm (simplified) 1: Input: k, n. Output: ˆH {H0, H1}. 2: For i = 1, 2, 3, . . . , C (log(k) + k ) do 3: Draw x1, . . . , xn D. 4: Define M 1 n Pn i=1 x i. 5: Define M := Ex N(0,I)[x i]. 6: If a=(i1, . . . , ji) such that |Ma M a|>d Cmλm 7: then Output H1 and terminate. 8: Return H0. The tester above is a simplified version. However, it is not fully correct, as we must ensure the concentration of the empirical tensor to bound the sample complexity. The Gaussian N(0, I) s empirical tensor is well-concentrated. While the parallel pancake s tensor might not concentrate well, this happens only when there is a Gaussian component much farther than d from the origin otherwise every sample from the parallel pancake is within O( d) in norm in high probability, and the empirical tensor is entrywise well-concentrated (e.g. by Hoeffding). This is also a testable condition: with log(k)/wmin samples, we will be able to check if every component is centered at most O( d) from the origin. The full version of the algorithm with this additional preliminary check, along with its correctness proof, are provided in Appendix D.1.
Open Source Code No The paper does not provide any statement about code availability or links to source code repositories.
Open Datasets No The paper is theoretical and does not describe experiments using specific datasets, thus no information on open datasets is provided.
Dataset Splits No The paper does not describe any experiments using specific datasets, therefore, it does not provide information on dataset splits.
Hardware Specification No The paper is theoretical and does not describe any experiments, therefore, it does not provide hardware specifications.
Software Dependencies No The paper is theoretical and does not describe any specific software implementations or dependencies with version numbers.
Experiment Setup No The paper is theoretical and does not describe any experiments, therefore, it does not provide details on experimental setup or hyperparameters.