[Submitted on 3 Jun 2026 (v1), last revised 29 Jul 2026 (this version, v3)]

View PDF HTML (experimental)

Abstract:Transformers consuming multi-channel scalar signals must embed $C$ simultaneous values into one $d_{\text{model}}$-dimensional vector per time step. We audit eight input encoders -- a shared-scalar baseline, per-channel linear projections, an orthogonality regulariser, a nonlinear MLP, block-partitioned concatenation, channel-independent and channel-as-token architectures, and a projected positional encoding -- on a synthetic benchmark where channel identity is informative and on ETTh1, scored by next-step negative log-likelihood. The headline is practical near-equivalence within a wide "top tier": the standard per-channel linear projection matches every alternative up to small, statistically real but practically modest differences. A direct geometric probe attributes this to a spontaneous orthogonalisation of the per-channel projections: with no explicit regulariser they tighten well beyond the near-orthogonality of random initialisation, letting the standard linear recover channel identity from the summed embedding. Two encoders lose decisively: the shared-scalar baseline collapses for information-theoretic reasons we make explicit, and the channel-independent PatchTST-spirit baseline overfits universally on the synthetic benchmark and underperforms on both. Paired tests resolve two small gaps: projecting the sinusoidal positional encoding through a learned linear layer edges the rest at small $C$ by extending this orthogonality to the positional subspace; a nonlinear MLP stem edges them at the largest $C$, with the gap shrinking under more training data. The practical recommendation: use the standard per-channel linear projection by default; reach for something more elaborate only when the task calls for it.

Submission history

From: Ossi Lehtinen [view email]
[v1] Wed, 3 Jun 2026 11:35:09 UTC (46 KB)
[v2] Mon, 8 Jun 2026 08:15:04 UTC (46 KB)
[v3] Wed, 29 Jul 2026 07:52:45 UTC (48 KB)