wifi-densepose/docs/research/sota-2026-05-22/R3_1-physics-informed-env-prediction.md

R3.1 — Physics-informed env_sig prediction at raw-CSI level: NEGATIVE (with a clear path forward)

Status: experimental result + scope correction · 2026-05-22

The plan

R3 (tick 12) showed MERIDIAN env-centroid subtraction recovers cross-room re-ID accuracy in the AETHER embedding space, but requires labelled examples in the new room. R3's "next research lever":

Use R6.1 forward operator + a coarse room map to PREDICT the env_sig without labelled examples — zero-shot transfer.

R6.1 (tick 18) shipped the multi-scatterer Fresnel forward operator. This tick implements the predicted-env approach at the raw CSI level (not the embedding level) and benchmarks it against R3's labelled MERIDIAN oracle.

Result

Two synthetic rooms (5×5 m diagonal link vs 4×6 m different link), 10 subjects with 0.85-1.15× body-size variation, 3 positions per room:

Configuration	1-shot K-NN accuracy
Within-room 1 baseline	100%
Within-room 2 baseline	100%
Cross-room raw (no env subtraction)	10% (= chance)
Cross-room labelled MERIDIAN (oracle)	10% (= chance)
Cross-room physics-informed env prediction	10% (= chance)

All three cross-room approaches collapse to chance. Not just the physics-informed one — even the labelled MERIDIAN oracle fails. This is meaningfully different from R3's tick-12 result where labelled MERIDIAN reached 100%.

Why R3 worked but R3.1 doesn't

R3 was simulated on a 128-dim AETHER-style embedding space where:

• person_signature, environment_signature, and noise were in independent random directions
• env_sig was a single fixed vector per room (no within-room positional variance)
• cosine normalisation partially absorbed the env shift

R3.1 is at the raw CSI level (52-dim complex) where:

• Subjects move to 3 positions per room — each position has its own complex CSI signature
• Per-position variance within a room can exceed per-subject variance between rooms
• Subtracting a single per-room centroid removes the mean position but not the variance

The headline gap: AETHER embedding space invariantises over within-room position; raw CSI does not. The cross-room problem at raw-CSI level is fundamentally harder than at the embedding level.

The honest takeaway

What R3 showed	What R3.1 shows
Cross-room re-ID works in embedding space with MERIDIAN	Cross-room re-ID doesn't work at raw-CSI level
Labelled centroid subtraction is enough	Labelled centroid subtraction is not enough at raw CSI
Physics-informed prediction is a worthwhile next step	Physics-informed prediction at raw-CSI level is also not enough

This is a third honest negative result for the loop (alongside R13 contactless BP and R12 NEGATIVE pre-PABS). The negative pattern: any cross-room method at raw-CSI level fails because position-variance is the dominant source of within-room CSI variation.

The path forward

The physics-informed env prediction approach is not dead — it just needs to be applied at the embedding level, not the raw-CSI level. The corrected architecture:

raw CSI → AETHER embedding head (position-invariant) → physics-informed env subtraction → cross-room K-NN

Or equivalently: subtract the physics-predicted env_sig from the AETHER head's output, not from the raw input. AETHER already does the heavy lifting of invariantising over position; the physics-informed prediction then has only the room-shift component to remove.

This requires AETHER (ADR-024) to be trained or fine-tuned, which is out of scope for this loop. The implementation roadmap is now clear:

1. AETHER head fine-tuned per-installation (ADR-024 baseline)
2. Physics-informed env_sig from R6.1 forward operator + room map
3. Subtract (2) from (1)'s output → invariantised embedding
4. K-NN matching across rooms with no labels in the new room

R3.1 says: the physics-informed prediction must be applied in the right space. The raw-CSI experiment exposes that the wrong space gives no lift.

Composes with prior threads

• R3 (cross-room re-ID) — R3.1 confirms R3's MERIDIAN-in-embedding-space result by showing the raw-CSI version fails. R3's choice to operate in embedding space was correct.
• R6.1 (multi-scatterer Fresnel) — provides the forward operator. R3.1 used it; the operator is correct; the application level was wrong.
• R12 PABS (POSITIVE) — operates on raw CSI directly but doesn't compare across rooms. PABS detects structural changes within a room; cross-room transfer needs an additional invariance layer (= AETHER).
• R14 / R15 / ADR-105 — the privacy framework still holds; AETHER + physics-env-prediction stays on-device per ADR-106.

Why this negative result is still useful

1. Surfaces an architecture error before implementation. Without this tick, a future engineer might attempt the obvious "subtract predicted env from raw CSI" approach and waste weeks. R3.1 documents that this fails.
2. Tightens the R3 implementation roadmap. The corrected architecture is now explicit.
3. Demonstrates the difference between embedding-space and raw-space approaches. This generalises beyond R3 — it informs every "subtract a learned/predicted nuisance" pattern in the codebase.

Honest scope

• 10 subjects with 0.85-1.15× body-size variation is a deliberately weak per-subject signature. Stronger biometric primitives (gait, breathing, RCS from R15) would give larger per-subject contrasts. The "raw CSI level fails" finding might be sensitive to this scale; with richer biometric input the raw-level approach might recover.
• The simulation uses 3 positions per room. With more positions (5-10), the failure would be sharper. With fewer (1), it would partially work.
• Position-variance dominance is geometry-specific. Long-narrow rooms vs square rooms have different ratios; this is one geometry.
• We didn't test "labelled MERIDIAN per-position-cluster" (cluster positions within a room, subtract per-cluster centroid). That might work for the labelled oracle; physics-informed equivalent would need a position-clustering layer.

What this DOES enable

• A negative result that prevents wasted implementation effort.
• A corrected architecture sketch: physics-informed env prediction at the embedding level (not raw level).
• A reference benchmark showing that the cross-room problem at raw-CSI level is genuinely hard, contextualising R3's embedding-level result.

What this DOES NOT enable

• The originally hoped-for zero-shot cross-room re-ID. That still needs the embedding-level implementation (R3.2, future).
• Any improvement to the existing within-room re-ID (which already works).
• Cross-installation re-ID — still prohibited by R3 + R14 + R15 + ADR-106.

What's next

• R3.2: embedding-level physics-informed env prediction (corrected architecture). Requires AETHER + R6.1 integration; out of scope for this loop.
• R12.1 (pose-PABS closed loop) — still the highest-leverage next implementation.
• ADR-107 (cross-installation federation) — still deferred.

Connection back

• R3 (POSITIVE in embedding space) — confirmed indirectly; raw-level failure shows why R3 operated at the embedding level.
• R6.1 — operator is correct; application level was wrong.
• R12 PABS (POSITIVE) — operates in raw space for structure detection (no cross-room transfer needed). PABS works at raw level because the comparison is within-room.
• R13 (NEGATIVE, physics floor) + R3.1 (NEGATIVE, architecture error) — two different kinds of negative result: one is a physics wall (R13), the other is a fixable design choice (R3.1).

Three kinds of negative result this loop has produced

This tick is the third honest negative — and the loop now has examples of all three categories:

1. R12 NEGATIVE → POSITIVE (revisited): missing tool (forward operator) blocked the right approach; tool became available later, approach worked.
2. R13 NEGATIVE → permanent: physics floor (5 dB shortfall) cannot be overcome by any tool; the negative is final.
3. R3.1 NEGATIVE → architecture-error: right idea, wrong application level; corrected architecture is now explicit but not yet implemented.

Knowing which category a negative result falls into is itself a research contribution. R3.1 sits in category 3.