Draft · v0.2 · Unreviewed
← Index/Project · Nadarasa ReductionDraft · v0.2 · Selene Emulator

Nadarasa · v0.4.2 · Selene experiment

G18 — Laplacian moments on a WL-blind graph pair

The quantum-TDA half of the SoftBank / Quantinuum white paper estimates spectral moments T_k = tr(Δ^k) of a graph Laplacian from a propagator trace rather than from a classical eigendecomposition. The sharpest small test case is a pair of graphs that 1-dimensional Weisfeiler–Leman colour refinement cannot distinguish: every vertex in both graphs has degree 2, so the refinement stabilises immediately and returns identical colourings. Their Laplacian spectra are different, and so are their moments from k = 3 upward.

Circuits
384
Shots / circuit
4096
χ²/dof
110.1
Max point
23.9σ

The pair

GraphV / EDegreesBetti (b₀, b₁)Laplacian spectrum
C66 / 62, 2, 2, 2, 2, 21, 10.0, 1.0, 1.0, 3.0, 3.0, 4.0
2C36 / 62, 2, 2, 2, 2, 22, 20.0, 0.0, 3.0, 3.0, 3.0, 3.0

Same vertex count, same edge count, same degree sequence — but C6 is connected with one independent cycle and 2C3 is two disjoint triangles. Topology sees it (b₀ = 1 vs 2); colour refinement does not.

What the circuit measures

The Laplacian is padded to 2³ = 8 dimensions, rescaled by 4 so the spectrum sits in [0, 1], and decomposed into 2022 Pauli strings. A Hadamard test on one ancilla plus three system qubits reads off ⟨x|e^(−iHτ)|x⟩ for each computational basis state x; averaging the eight of them gives the normalised trace f(τ) = tr(e^(−iHτ)) / 8. Real and imaginary parts come from the same circuit with and without the ancilla S† rotation. The propagator is Trotterised at 3 steps, which costs up to 1.6e-2 in trace error at the longest τ — about two point-σ, so the measured curves track the exact ones to a couple of percent rather than to shot noise. That systematic is per-graph and largely common-mode: the measured differences below still track the exact differences closely.

Deterministically sweeping all eight basis states beats a random purification here: at three qubits it is only 8 circuits per (τ, part), and it removes the state-preparation variance entirely.

Model-free separation: the trace curves themselves

Before fitting anything, the two measured curves can simply be subtracted. Each averaged point carries binomial error 0.0078, so the difference is a clean χ² test with 24 degrees of freedom.

τRe C6Re 2C3Δ measuredΔ exactσ
0.500.96660.96470.00190.00010.2
0.860.89950.9008-0.00130.00100.2
1.230.81270.79960.01310.00411.7
1.590.69620.68510.01110.01081.4
1.950.58360.54800.03560.02284.6
2.320.46250.41660.04590.04105.9
2.680.34640.29490.05150.06556.6
3.050.26510.18410.08100.095110.4
3.410.21230.07080.14150.127518.1
3.770.17830.02070.15760.158820.2
4.140.1768-0.00960.18640.184523.9
4.500.19170.01130.18040.199723.1

χ² = 2643 over 24 points = 110.1 per degree of freedom, with a single point reaching 23.9σ. The emulator distinguishes the pair decisively, using a quantity 1-WL cannot see.

Recovered moments

Fitting the truncated Taylor series f(τ) = Σ (−iτ)^k m_k / k! converts the curves into moments. Two truncations are shown: order 8 (near-unbiased, high variance) and order 6 (low variance, but truncation pushes weight from T₃ into T₄).

GraphkExact T_kFit (order 8)Fit (order 6)
C611212.0 ± 0.111.5 ± 0.1
23635.0 ± 0.634.9 ± 0.3
3120118.0 ± 4.098.3 ± 1.4
4420366.9 ± 32.4360.8 ± 9.6
2C311212.2 ± 0.111.8 ± 0.1
23636.6 ± 0.635.3 ± 0.3
3108112.7 ± 4.099.0 ± 1.4
4324367.4 ± 32.4291.2 ± 9.6

Moment-by-moment separation

kExact ΔMeasured Δ (order 8)σMeasured Δ (order 6)σ
100.10.7σ0.32.4σ
201.61.8σ0.40.8σ
3125.30.9σ0.80.4σ
4960.50.0σ69.65.1σ

T₁ and T₂ agree between the two graphs, as they must — those moments are fixed by the vertex and edge counts, which the pair shares. The separation appears only at higher order. At order 8 the T₃ gap (12) sits inside its own error bar, so on this shot budget the per-moment test is not conclusive; at order 6 the truncation reassigns the discrepancy to T₄, where it shows up at 5.1σ. The honest summary: the moment vector separates the pair, but attributing the separation to one specific k needs more shots.

Caveats

  • Noiseless Quest. G17 is the template for what an H2-class error ladder would do to these curves; that ladder has not been run here.
  • Trotter error is not negligible here. At 3 steps it reaches 1.6e-2 in the trace, comparable to two point-σ, and it accounts for part of the gap between the order-8 moment fits and the exact values. More steps would fix it at proportionally more circuit depth.
  • Three system qubits, six-vertex graphs. This is a mechanism check, not a scaling claim — the interesting regime is where the Laplacian no longer fits in a classical eigensolver.
  • High moments are dominated by fit variance, not by the emulator. The Taylor-fit conditioning (fit order and τ grid) dominates the error far more than the shot count does; doubling the order roughly quadruples σ(T₃).
  • Results are committed static JSON. Nothing quantum runs at request time.

Reproduce

pip install --target .pydeps -r quantum/requirements.txt
PYTHONPATH=.pydeps python -m quantum.tda.sweep

Backend: selene-sim Quest (noiseless). Each (graph, τ, basis state, part) circuit caches under quantum/tda/_cache_tda/, so the 384-circuit sweep is resumable across sandbox sessions.