LIT-181: Beyond the Birkhoff Polytope: Spectral-Sphere-Constrained Hyper-Connections

Liu et al. (2026) — ARXIV-2603.20896

Key takeaways

• Accepts mHC’s diagnosis — unconstrained mixing breaks the identity mapping and destabilises training — and attacks its remedy, naming three limitations of restricting the residual matrix to the Birkhoff polytope:
  1. Identity degeneration — the learned matrices collapse toward the identity, so the cross-stream interaction the extra width was for diminishes.
  2. An expressivity bottleneck — non-negativity forbids subtractive feature disentanglement. A stream can only ever be a positive blend of the others.
  3. Parameterization inefficiency — unstable Sinkhorn iterations, or the factorial scaling of permutation-based alternatives.
• sHC shifts the feasible set from a rigid polytope to a spectral-norm sphere, which permits negative entries and so unlocks subtractive interactions, while still bounding the mixing. It removes the Sinkhorn projection and the factorial parameterization together.

Standing in the anthology

The second independent critique of LIT-140, and the pair is now the most interesting thing in the record’s residual chain. LIT-151’s oHC says the doubly stochastic set is bounded above but not below, so the mixing can only shrink what distinguishes the streams; this says the same set is degenerate near the identity and cannot subtract at all. Different arguments, same verdict, and opposite remedies — the rotation group SO(n) versus a spectral sphere.

Both critique a practice the record recommends. SOTA-136 is Proposed and stays that way; what has changed since it was filed is that its constraint is now the least well-defended part of the design, disputed by two groups within a month, while the thing all four papers agree on — widen the residual stream, constrain the mixing somehow — is untouched.

That agreement is now SOTA-169, Proposed, filed three days after this note and sourcing it alongside LIT-141, LIT-140, LIT-151 and LIT-152. Version 1 of this section asked for it as something “the record should eventually carry”, and that sentence stayed standing for eleven days after it was answered — a request written in prose carries no code, resolves to nothing, and so nothing retracts it when it is met.

The line this is a step in is on the lines-of-work page, walked from the extends: field rather than restated here.