Cekirge Perturbation Report v4
Abstract
This report introduces a deterministic σ-regularized framework for transformer-style models that computes attention and decoding in closed form without iterative gradient descent. By constructing nonsingular query (Q), key (K), value (V), and output (W) matrices, the method ensures invertibility, numerical stability, and reproducibility. Stability is probed by uniform matrix perturbations of amplitude ε, and the resulting change in loss ΔL = L(ε) − L(0) is recorded. Empirically, ΔL increases nearly linearly with a gently decreasing slope (≈1.84 → ≈1.70 for ε = 0.01 → 0.10), confirming energy-bounded forward determinism. This linear ε–ΔL dependence distinguishes the Çekirge method from stochastic training, revealing a predictable, non-chaotic energy profile. The deterministic σ-regularized formulation provides a foundation for algebraic pretraining, analytical model verification, and thermodynamically aligned, energy-efficient artificial intelligence.
Full text
Cekirge Perturbation Report v4 '2,]HQRGR RESTRICTED ɐǦǣ ǤǤȄȋȌǡ ʹͲʹͷ ďƐƚƌĂĐƚ ɂȂȟ.ɐǦǡ ǦǦǤȀȀ ɂȟ ǡǦǤ .Ǧ ǡǤ ϭDĞƚŚŽĚƐ ȂǤ ɐǦǤ ȋͲȌɂאȏͲǡ ͲǤͳͲȐȀȀȋȌǤȟȋɂȌαȋɂȌΫȋͲȌ Ǥ ϮŽƌĞƋƵĂƚŝŽŶƐ;ZĞĂĚĂďůĞ&ŽƌŵͿ ǦȋǦȌǣڅαȋᤘΪɐȌ·ͽᤘ ǣɁαΫǡᤘɁǤ ǣȟȋɂȌαȋΪɂȟǡΪɂȟǡΪɂȟȌΫȋǡǡȌǤ https://orcid.org/0000-0001-8075-2306
ϯZĞƐƵůƚƐȴ>ʹɸ^ƚĂďŝůŝƚLJƵƌǀĞ ͳȄ.ȋȟȂɂȌǤǡȟɂ ɐǦǦǤ ϰ^ĞŶƐŝƚŝǀŝƚLJdĂďůĞ ɂȋȌ ȟȋȌ ȟȀɂȋȌ ͲǤͲͲͲ ͲǤͲͲͲͲͲͲ ͲǤͲͳͲ ͲǤͲͳͺͶͶ͵ ͳǤͺͶͶ ͲǤͲʹͲ ͲǤͲ͵ͷ͵ͺ ͳǤͺʹ ͲǤͲ͵Ͳ ͲǤͲͷͶʹͻͷ ͳǤͺͳͲ ͲǤͲͶͲ ͲǤͲͳʹ͵ ͳǤͻ͵ ͲǤͲͷͲ ͲǤͲͺͺͺ͵ͳ ͳǤ ͲǤͲͲ ͲǤͳͲͷʹͺ ͳǤͲ ͲǤͲͲ ͲǤͳʹʹͳʹ͵ ͳǤͶͷ ͲǤͲͺͲ ͲǤͳ͵ͺ͵ʹ͵ ͳǤʹͻ ͲǤͲͻͲ ͲǤͳͷͶʹ͵ ͳǤͳͶ ͲǤͳͲͲ ͲǤͳͻͺͲ ͳǤͻͻ γͳǤͺͶɂαͲǤͲͳγͳǤͲɂαͲǤͳͲǡǦ Ǥ
ϱŝƐĐƵƐƐŝŽŶ .ɐǦǡ Ǥ ǡǤȟȂɂ .ǡǡ Ǥ ϲZĞĨĞƌĞŶĐĞƐ;^ĞůĞĐƚĞĚͿ ȏͳȐǦǤ ȏʹȐǤ ȏ͵ȐǦǤ ϳdŚĞDĞĂŶŝŶŐŽĨƚŚĞ'ŽůĚĞƚĞƌŵŝŶŝƐƚŝĐ>ĂǁŽĨ>ĞĂƌŶŝŶŐ Ǯǯ.Ǧ ǡɐǦǤαȋᤘΪɐȌ·ͽᤘ ǡǤȟȂɂ Ǧǡ Ǥǡǡǡǡ Ȅ Ǥǣǡ Ǥ ϴ^ƵďŵŝƐƐŝŽŶ^ƚĂƚĞŵĞŶƚ ǡǦǤ ɐǦǦǤ ǣǤǤ.ǡȋȌǡ Ǥ ǣʹͲʹͷǦͳͲǦͳͻ ǣȋȀȌ