Skip to content

Dynamic Laplacian on SH substrate — PR #335 revisited and CONFIRMED (2026-05-12)

Origin: PR C of 3-PR sequence (Zernike #338 → PCA canonical #339 → this). PR #335 falsified the dynamic-Laplacian iteration on voxel substrate because the active region geometrically disconnected (positive and negative Fiedler regions formed separate blobs). User question: does it work on a connected substrate?

Verdict: WIN. On a globally-connected sphere substrate, the dynamic Laplacian iteration converges in 4 steps to a non-trivial fixed point (correlation 0.23 with initial — structurally distinct). PR #335 was substrate-limited, not concept-limited. The user's original architectural intuition (fiber feedback to graph structure) was correct.

Reproduce: python -X utf8 docs/srmech/notes/dynamic_laplacian_sh_substrate_script.py. Runtime ~3s.


Setup — same iteration as PR #335, different substrate

Element PR #335 (FALSIFIED) PR C (this)
Substrate 32³ voxel grid, field
Active region 4621 voxels in 2 disconnected components 7200 nodes, all reachable
Connectivity DISCONNECTED (2532 + 2089 = 4621) CONNECTED (BFS reaches all 7200)
Edge weights w = cos²(Δθ/2) — same w = cos²(Δθ/2) — same
Update rule Fiedler-then-phase — same Fiedler-then-phase — same
Sign alignment to previous Fiedler — same to previous Fiedler — same

The only substantive change is the substrate — voxel grid → sphere surface. Everything else identical to PR #335.


Connectivity diagnostic — the load-bearing difference

Check PR #335 (voxel) PR C (sphere)
BFS-reachable nodes 2,532 / 4,621 7,200 / 7,200
Unreachable nodes 2,089 0
Graph topology 2 disconnected components 1 connected component
Matches positive/negative bipartition? YES (2532 ≈ +; 2089 ≈ −) N/A (single component)

PR #335's voxel substrate baked the bipartition into geometric disconnection. The sphere substrate doesn't.


Iteration results

| Step | λ_Fiedler | ||Δphase|| | Correlation with ψ₀ | |---|---:|---:|---:| | 0 | 0.00274 | 111.39 | 0.228 | | 1 | 0.00273 | 0.192 | 0.228 | | 2 | 0.00273 | 0.001 | 0.228 | | 3 | 0.00273 | 0.00001 | 0.228 |

Converged at step 3 (||Δphase|| = 10⁻⁵, well below tolerance 10⁻⁴).

Final correlation with initial: 0.228 — significantly different from 1.0. The converged phase field ψ* is structurally distinct from the initial ψ₀.

λ_Fiedler ≈ 0.003 across iterations — small but non-zero. The dynamic Laplacian is connected (kernel dim 1, the constant eigenvector). The Fiedler eigenvector is well-defined.


Outcome classification

non_trivial_fixed_point — confirmed after classifier fix.

(The first run's classifier bug — using max over last 3 steps instead of final step — mislabeled this as slow_or_undecided. Fixed: classify by final-step behavior. Both behaviors describe the same underlying iteration; only the label changed.)


What this confirms about PR #335's question-tree branch

The user's original PR #335 architectural question was:

"Now that we can draw our hypervector as voxels, does that mean we could also use fiber content to 'reshape' voxels for dynamic graph-Laplacian things?"

PR #335 answered "no, falsified" on the voxel substrate. The vacuousness was geometric: the bipartition was already encoded in voxel topology before iteration started.

This PR answers "yes, with connected substrate." The user's architectural intuition was sound; the previous spike's negative result was substrate-specific, not concept-fatal.

Refined claim: dynamic-Laplacian iteration via fiber phase coherence works on substrates where the graph topology doesn't pre-bake the field's bipartition. Sphere surface qualifies; thresholded voxel active-region doesn't.


Architectural lesson — substrate is part of the primitive

A "dynamic Laplacian" primitive isn't fully specified by the iteration rule. It's specified by (iteration rule) × (substrate). The substrate choice can:

  1. Pre-encode the data into geometry (PR #335 voxel case) — primitive is vacuous
  2. Stay connected through data variation (this PR sphere case) — primitive does real work
  3. (Hypothetically) Provide its own geometric structure that conflicts with data — primitive may diverge

For the srmech §4.2 catalog: the "dynamic state-dependent operator" entry should specify both the rule AND the required substrate properties (connectivity, non-bipartition-aligned topology, etc.).


Connection to MFO §VII.4.1.1 (Hopf bundle, PR #332)

PR #332's continuum framework predicts dynamic-Laplacian behavior on a principal U(1)-bundle. The sphere surface here is the base S² of a (trivial-by-default) U(1)-bundle. The dynamic Laplacian discovers a fixed point that lives on this S² substrate — a discrete analog of finding harmonic forms on a continuous principal-bundle base.

To extend: instead of trivial U(1)-bundle (just S²), use a non-trivial bundle with Chern class 1 — i.e., the actual Hopf bundle S³ → S². The dynamic Laplacian on the full S³ total space would test what PR #332's continuum framework predicts: that fibre harmonics carry orientation-encoding information.

This is a natural follow-up spike.


Updated question-tree status

PR Question Answer
#333 Voxel HDC + Hodge Voxel HDC works; Hodge β_k separates fold classes
#334 Universal eigenvalue statistics? Proteins are GOE-universal (Maass class)
#335 Reshape voxels via dynamic L? Falsified on voxel substrate (geometric disconnect)
#336 Dynamic metric vs static L Reveals rotation breaks voxel HDC
#337 SH power for SO(3) invariance Works but discrimination poor
#338 3D Zernike for more discrim Falsified (collapsed)
#339 PCA canonical-frame WIN (best geometric fingerprint)
#340 (this) Dynamic L on connected substrate? WIN (non-trivial fixed point)

What this opens up

  1. Dynamic Laplacian on actual Hopf bundle — non-trivial U(1) bundle over S². Tests MFO §VII.4.1.1 continuum framework directly.
  2. Dynamic Laplacian on PCA-canonical sphere — combine PR #339's canonical orientation with this PR's connected substrate. Best of both: SO(3) invariance + non-trivial fixed point.
  3. Convergence analysis — what does the fixed-point ψ* actually look like spatially? Does it correspond to a known structural feature of ubiquitin?
  4. Cross-protein test — does 1BPI converge to the SAME fixed point as 1UBQ (would indicate a sphere-universal mode, not a protein-specific one)?

These are flagged for future spikes, not in scope for PR C.


Honest caveats

  • Substrate-aware verdict: the dynamic Laplacian works on the sphere substrate at the specific (60×120) lat-lon grid tested. Coarser grids might degenerate; finer grids might give the same answer at higher cost.
  • Initial correlation 0.23: the converged ψ* is distinct from ψ₀ but not maximally distinct. Whether 0.23 is a "deep" non-triviality or a near-orthogonal one depends on downstream interpretation.
  • λ_Fiedler ≈ 0.003: small. The dynamic Laplacian has spectral structure but it's compressed near the kernel. Larger lat-lon grids might give sharper spectral separation.
  • Classifier was buggy on first run (max-over-window heuristic). Fixed to use final-step behavior. Both runs describe the same iteration.

Files

  • dynamic_laplacian_sh_substrate_script.py — reproducible spike (~3s)
  • dynamic-laplacian-sh-substrate-per-step-2026-05-12.ndjson — per-step + outcome
  • dynamic-laplacian-sh-substrate-trajectory-2026-05-12.png — convergence plot + initial/final phase fields on sphere
  • dynamic-laplacian-sh-substrate-2026-05-12.md — these findings

Citations

  • PR #335 falsification of voxel-substrate dynamic Laplacian
  • MFO §VII.4.1.1 (PR #332) continuum framework
  • Lat-lon sphere discretization: standard in geophysics / climate modeling