The Algorithmic Phase Transition for Correlated Spiked Models

Modern multi-modal learning often relies on the premise that jointly analyzing multiple, related datasets can yield more powerful inferences than processing each one in isolation. We study this through the lens of a pair of spiked random matrices with correlated spikes. By proposing a novel subgraph counts algorithm, we show that the correlation between the spikes can be exploited for inference even in certain regimes where inference in each individual matrix is believed to be computationally intractable. Furthermore, we provide evidence for a matching computational lower bound based on the low-degree polynomial framework, suggesting our algorithm is optimal. Our results thus establish a new computational phase transition in correlated spiked models, delineating the boundary between what is efficiently possible and what is not. Based on arXiv:2511.06040.

The talk will be streamed through "Tencent": https://meeting.tencent.com/dm/4TGPAsgr42lx.
meeting code: 865-161-635
To join via a web browser: go to https://voovmeeting.com/; click "Log in" in the upper-right corner and login either by a Google/Apple account, or by a verification code received on phone or email; click "Join now" in the upper-right corner and type the number of the meeting 865-161-635 in the "Enter meeting ID" field; when prompted, select "Join from browser".