{"id":"arxiv/2609.09556","url":"https://arxiv.org/abs/2609.09556","tags":["stat.ML","cs.AI","cs.IR","cs.LG","math.PR"],"score":0,"title":"High-probability guarantees for linear accessibility in feature superposition","domain":"stat.ML","source":"arxiv","authors":["Enrico Vompa"],"reviews":0,"subject":"stat.ML","summary":"Neural networks can leverage feature superposition to encode more concepts than dimensions, but cross-feature interference constrains the linear accessibility of simultaneously active features. By framing linear accessibility as a compressed sensing problem, we derive high-probability bounds for fixed supports under subgaussian noise, proving the sufficient dimension scales linearly ($d=O_{\\varepsilon}(k \\log m)$) rather than prior worst-case quadratic limits. We characterize the asymmetry between active and inactive interference and the trade-off between interference and observation-noise…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-09-09","reactions":{}}