PSGD-QEQ equals KL-AIRM up to step size, author corrects gradient uniqueness claim
YouJiacheng · x · 2026-09-10
The author corrects an earlier claim: requiring the gradient to depend linearly on Σ makes the Euclidean gradient unique up to a change of basis, so KL and PSGD objectives yield identical updates under the same geometry. A further derivation shows the gradient isn't truly unique but coincidentally offset by geometry differences — since dp=d(q²)=2qdq=-2q²r=-2pr, PSGD-QEQ equals KL-AIRM up to a step size.
Related event: Correction: PSGD and KL-AIRM Gradients Differ Only by a Step Size(2 posts)→
More from Research
- Perturb the Physics, Not the Network: Magnetic Impurities Unlock Hamiltonian Learning — bravo_abad · 2026-09-10
- Vision-force fusion robot dressing handles moving arms: 85% arm coverage across 264 real trials — stepjamUK · 2026-09-10
- The J-lens Explained: Reading and Rewriting LLMs' Unspoken Concepts — CatAstro_Piyush · 2026-09-10
- Group Bench: ~100 group theory problems to benchmark your AI agents, with a dated progress map — Sauers_ · 2026-09-10
- Group Bench launches ~100 group theory problems for testing AI agents — Sauers_ · 2026-09-10
- ELLIS PhD Program Opens 2026 Applications With Cross-Border Co-Supervision — ArthurGretton · 2026-09-10