Alexander Johansen

Anupama Sridhar, Alexander Johansen

Long chain-of-thought reasoning and agentic tool-calling produce traces spanning tens of thousands of tokens, yet Transformer KV caches grow linearly with sequence length, creating a memory bottleneck on commodity hardware. State-space models offer constant-memory recurrence but suffer a memory cliff: retrieval accuracy collapses once the gap between a stored fact and its query exceeds the effective horizon of the recurrent state. We introduce Echo, a KV-cache-free associative recall architecture built around Spectral Koopman Attention (SKA); a drop-in replacement for attention layers that augments SSM blocks with a closed-form dynamical operator whose sufficient statistics are accumulated in constant memory with no KV cache. Echo fits a spectral linear system to the key and value history via kernel ridge regression and retrieves through a learned power-iterated filter, all from $O(r^{2})$ streaming state where $r$ is a small projection rank. On the Multi-Query Associative Recall benchmark, a pure Mamba-2 SSM fails to exceed chance accuracy (${\sim}3\%$) across all gap lengths and KV-pair counts, while at the 50M parameter scale SKA-augmented models achieve $100\%$ retrieval accuracy on every configuration tested, including distractor gaps of $4{,}096$ tokens with $32$ KV pairs. Across five additional transfer benchmarks including needle-in-a-haystack, tool-trace, and multi-hop retrieval, SKA consistently outperforms both pure SSM and SSM+Attention hybrids while maintaining constant inference memory. Ablations confirm that the spectral operator, not the prefix masking strategy, drives the retrieval gain.

Anupama Sridhar, Alexander Johansen

First-order adaptive optimization methods like Adam are the default choices for training modern deep neural networks. Despite their empirical success, the theoretical understanding of these methods in non-smooth settings, particularly in Deep ReLU networks, remains limited. ReLU activations create exponentially many region boundaries where standard smoothness assumptions break down. \textbf{We derive the first \(\tilde{O}\!\bigl(\sqrt{d_{\mathrm{eff}}/n}\bigr)\) generalization bound for Adam in Deep ReLU networks and the first global-optimal convergence for Adam in the non smooth, non convex relu landscape without a global PL or convexity assumption.} Our analysis is based on stratified Morse theory and novel results in Kakeya sets. We develop a multi-layer refinement framework that progressively tightens bounds on region crossings. We prove that the number of region crossings collapses from exponential to near-linear in the effective dimension. Using a Kakeya based method, we give a tighter generalization bound than PAC-Bayes approaches and showcase convergence using a mild uniform low barrier assumption.

Anupama Sridhar, Alexander Johansen

Temporal Difference Learning (TD(0)) is fundamental in reinforcement learning, yet its finite-sample behavior under non-i.i.d. data and nonlinear approximation remains unknown. We provide the first high-probability, finite-sample analysis of vanilla TD(0) on polynomially mixing Markov data, assuming only Holder continuity and bounded generalized gradients. This breaks with previous work, which often requires subsampling, projections, or instance-dependent step-sizes. Concretely, for mixing exponent $β> 1$, Holder continuity exponent $γ$, and step-size decay rate $η\in (1/2, 1]$, we show that, with high probability, \[ \| θ_t - θ^* \| \leq C(β, γ, η)\, t^{-β/2} + C'(γ, η)\, t^{-ηγ} \] after $t = \mathcal{O}(1/\varepsilon^2)$ iterations. These bounds match the known i.i.d. rates and hold even when initialization is nonstationary. Central to our proof is a novel discrete-time coupling that bypasses geometric ergodicity, yielding the first such guarantee for nonlinear TD(0) under realistic mixing.