{"paper":{"title":"FaSST: Fast Sparsifying Secondary Transform","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"FaSST factorizes data-driven secondary transforms into mode-adaptive Givens rotations to match LFNST performance at far lower complexity.","cross_cats":["eess.SP"],"primary_cat":"eess.IV","authors_text":"Antonio Ortega, Darukeesan Pakiyarajah, Debargha Mukherjee, Eduardo Pavez, Samuel Fern\\'andez-Mendui\\~na","submitted_at":"2026-05-14T17:08:16Z","abstract_excerpt":"Data-dependent secondary transforms, which aim to decorrelate coefficients of a separable primary transform, can improve residual coding efficiency; however, their deployment is often constrained by computational complexity. Recent video codecs use variants of the low-frequency non-separable transform (LFNST), which discards some high-frequency secondary transform coefficients, limiting achievable coding gains. Moreover, existing data-dependent secondary transforms lack explicit rate-distortion (RD) optimal design criteria. In this work, we propose a framework for designing low-complexity data"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Mode-adaptive FaSST matches the RD performance of LFNST while reducing the number of computations by 83.67%. Moreover, by avoiding fixed-coefficient truncation, FaSST achieves up to 1.80% BD-rate savings relative to LFNST while operating at 66.24% lower complexity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the alternating-minimization procedure combined with approximate Givens factorization produces transforms whose rate-distortion performance remains close to the original data-driven SOTs across the full range of intra prediction modes and content types used in AV2 testing.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FaSST approximates sparse orthonormal transforms with mode-adaptive Givens rotation sequences to produce low-complexity secondary transforms for AV2 intra residuals that match LFNST rate-distortion performance at 83.67% lower complexity and deliver up to 1.80% BD-rate savings.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FaSST factorizes data-driven secondary transforms into mode-adaptive Givens rotations to match LFNST performance at far lower complexity.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"9d04983890736f6c84e8b21bfd60f00facf8c6e9981956fbaa99798245f39f96"},"source":{"id":"2605.15086","kind":"arxiv","version":1},"verdict":{"id":"5faa7d63-ed01-46d1-93fa-4c4686ae032f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T03:14:30.967662Z","strongest_claim":"Mode-adaptive FaSST matches the RD performance of LFNST while reducing the number of computations by 83.67%. Moreover, by avoiding fixed-coefficient truncation, FaSST achieves up to 1.80% BD-rate savings relative to LFNST while operating at 66.24% lower complexity.","one_line_summary":"FaSST approximates sparse orthonormal transforms with mode-adaptive Givens rotation sequences to produce low-complexity secondary transforms for AV2 intra residuals that match LFNST rate-distortion performance at 83.67% lower complexity and deliver up to 1.80% BD-rate savings.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the alternating-minimization procedure combined with approximate Givens factorization produces transforms whose rate-distortion performance remains close to the original data-driven SOTs across the full range of intra prediction modes and content types used in AV2 testing.","pith_extraction_headline":"FaSST factorizes data-driven secondary transforms into mode-adaptive Givens rotations to match LFNST performance at far lower complexity."},"references":{"count":31,"sample":[{"doi":"","year":2026,"title":"FaSST: Fast Sparsifying Secondary Transform","work_id":"6485b07b-2706-4655-b1ba-080584dd46b6","ref_index":1,"cited_arxiv_id":"2605.15086","is_internal_anchor":true},{"doi":"","year":null,"title":"coefficient dropping","work_id":"daa4b745-907a-46a8-a1e7-8be3268033a8","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Moreover, KLTs and their approx- imations do not account for quantization effects","work_id":"32914eb2-603e-4dc4-997e-77857707104a","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"given the coefficients{ ˆyi}me i=1, we optimize the transform: min SJ meX i=1 ∥ˆxi −S J ˆyi∥2 2 s.t.S J = JY j=1 G(mj, nj, θj).(6) It can be shown that the above problem is equivalent to max SJ tr(ΓSJ","work_id":"2a6613c9-a683-432d-bf02-6fb652829533","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Specifically, images from the CLIC dataset [20] are compressed using A V2, and residual blocks of sizes 8×8,16×16, and32×32are collected","work_id":"f1d6bab2-57ce-4e98-9619-0aa7ba28d155","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":31,"snapshot_sha256":"2f10db13debd16dbfe3be60d95cebdb9199a07adaded0922e528887e3888c443","internal_anchors":2},"formal_canon":{"evidence_count":1,"snapshot_sha256":"5a2efaceeaa4261c92f30b97800c9d4a7be0ac38f6d969ee960551c4a30b75df"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}