repro-ssm-duality / logbook.json
snaykey's picture
logbook.json (verbatim titles)
8a54a40 verified
Raw
History Blame Contribute Delete
2.72 kB
{
"schema_version": 1,
"title": "Reproduction: On Structured State-Space Duality",
"emoji": "๐Ÿ”",
"space_id": "snaykey/repro-ssm-duality",
"paper": {
"arxiv_id": "2510.04944",
"openreview_id": "DKathyl3XN"
},
"tags": [
"icml2026-repro",
"paper-DKathyl3XN"
],
"updated_at": "2026-07-28T00:00:00+00:00",
"root": {
"slug": "index",
"title": "Reproduction: On Structured State-Space Duality",
"file": "pages/index.md",
"children": [
{
"slug": "executive-summary",
"title": "Executive summary",
"file": "pages/executive-summary/page.md",
"children": []
},
{
"slug": "claim-1-prop-3-1-scalar-identity",
"title": "Proposition 3.1 establishes the scalar-identity case, showing SSMs with state matrices A_t = a_t I_N are equivalent to 1-semiseparable masked attention with a causal mask (Section 3.2).",
"file": "pages/claim-1-prop-3-1-scalar-identity/page.md",
"children": []
},
{
"slug": "claim-2-thm-4-1-necessary-sufficient",
"title": "Theorem 4.1 gives a necessary and sufficient condition for an N-dimensional SSM to admit a 1-semiseparable masked-attention dual: its corresponding N-semiseparable matrix must have diagonal blocks each containing at most N new columns (Section 4.4, Definition 4.2).",
"file": "pages/claim-2-thm-4-1-necessary-sufficient/page.md",
"children": []
},
{
"slug": "claim-3-diagonal-ssd-complexity",
"title": "General diagonal SSMs are shown to preserve the same O(NTd) training complexity as the scalar-identity case (Algorithm 1) while supporting richer, per-channel dynamics (Section 4.3).",
"file": "pages/claim-3-diagonal-ssd-complexity/page.md",
"children": []
},
{
"slug": "claim-4-prop-4-1-nss-nsss",
"title": "Proposition 4.1 proves the equivalence between N-semiseparable matrices and N-SSS (sequentially semiseparable) representable matrices (Section 4.4).",
"file": "pages/claim-4-prop-4-1-nss-nsss/page.md",
"children": []
},
{
"slug": "claim-5-sec-5-softmax-general-ssm",
"title": "Section 5 demonstrates a limitation: softmax attention is incompatible with the semiseparable-matrix duality framework due to rank explosion, and general (non-diagonal) SSMs may lack any attention dual despite having low state dimension (Section 5).",
"file": "pages/claim-5-sec-5-softmax-general-ssm/page.md",
"children": []
},
{
"slug": "conclusion",
"title": "Conclusion",
"file": "pages/conclusion/page.md",
"children": []
}
]
}
}