wai-quantum 0.3.33: noisy circuits, and where the digital error model is off #156

Merged

dcharlot merged 2 commits from wai-quantum-noisy into main 2026-10-05 11:03:17 -04:00

Conversation 0 Commits 2 Files changed 13 +839 -9

dcharlot commented 2026-10-05 08:08:11 -04:00

Owner

Copy link

Stacked on #155 (→ #151). Merge those first; this diff then reduces to the one commit.

What

quantum_noisy (wai-quantum 0.3.33) simulates noisy circuits and checks the fidelity arithmetic that quantum-advantage claims rest on.

• Trajectories on the state vector. Each trajectory applies, after every gate, one randomly chosen branch of each noise channel on the qubits that gate touched:
  • depolarizing and dephasing noise, which stay unitary;
  • amplitude damping, renormalised after each jump;
  • readout flips on the measured bits.
• An exact density-matrix referee up to 12 qubits. It stores ρ as a state on 2n qubits, applies U ⊗ U* per gate and Kraus sums per channel, and reuses the state-vector kernels.
• Fidelity measures. Cross-entropy fidelity (XEB) in four forms: sampled, exact, the noiseless circuit's own, and the digital error model's prediction.
• CLI. wai-quantum noisy ….

Checked

• Expectations. 4,000 trajectories land within four standard errors of the exact values, with every noise channel on.
• Samples. Sampled bitstrings follow the exact distribution, including readout flips.
• XEB. The sampled XEB converges to the exact value.
• Determinism. Runs are identical at any thread count. The new pin holds natively and under wasm32-wasip2, and all 18 pins pass. CLI output is byte-identical between the two, threaded or not.
• Lint and tests. Clippy is clean on the new code (two type aliases added), the prose checks pass, and all 415 library tests pass. The two heaviest new tests were trimmed to about 25 s each in debug, which is how CI runs them.

Finding: the digital error model understates fidelity

On 8-qubit brickwork circuits under depolarizing noise, the exact XEB (normalised by the noiseless circuit's own) sits above the digital model's prediction:

Two-qubit error rate	Gap over 4 to 48 layers
0.005	3.5–4.5%
0.01	7–10%

The gap does not grow with depth, so it isn't a per-gate bias. It behaves like a fixed set of errors near the measurement that the model counts as destroying the signal when they don't. A Z error right before a computational-basis measurement, for example, changes no outcome at all.

On a circuit that does not scramble, the gap is wider: 0.71 exact against 0.60 predicted.

The tables are in the module documentation, and a regression test holds the depth-independence.

🤖 Generated with Claude Code

Stacked on #155 (→ #151). Merge those first; this diff then reduces to the one commit. ## What `quantum_noisy` (wai-quantum 0.3.33) simulates noisy circuits and checks the fidelity arithmetic that quantum-advantage claims rest on. - **Trajectories on the state vector.** Each trajectory applies, after every gate, one randomly chosen branch of each noise channel on the qubits that gate touched: - depolarizing and dephasing noise, which stay unitary; - amplitude damping, renormalised after each jump; - readout flips on the measured bits. - **An exact density-matrix referee** up to 12 qubits. It stores ρ as a state on 2n qubits, applies `U ⊗ U*` per gate and Kraus sums per channel, and reuses the state-vector kernels. - **Fidelity measures.** Cross-entropy fidelity (XEB) in four forms: sampled, exact, the noiseless circuit's own, and the digital error model's prediction. - **CLI.** `wai-quantum noisy …`. ## Checked - **Expectations.** 4,000 trajectories land within four standard errors of the exact values, with every noise channel on. - **Samples.** Sampled bitstrings follow the exact distribution, including readout flips. - **XEB.** The sampled XEB converges to the exact value. - **Determinism.** Runs are identical at any thread count. The new pin holds natively and under wasm32-wasip2, and all 18 pins pass. CLI output is byte-identical between the two, threaded or not. - **Lint and tests.** Clippy is clean on the new code (two type aliases added), the prose checks pass, and all 415 library tests pass. The two heaviest new tests were trimmed to about 25 s each in debug, which is how CI runs them. ## Finding: the digital error model understates fidelity On 8-qubit brickwork circuits under depolarizing noise, the exact XEB (normalised by the noiseless circuit's own) sits above the digital model's prediction: | Two-qubit error rate | Gap over 4 to 48 layers | |---|---| | 0.005 | 3.5–4.5% | | 0.01 | 7–10% | The gap does not grow with depth, so it isn't a per-gate bias. It behaves like a fixed set of errors near the measurement that the model counts as destroying the signal when they don't. A Z error right before a computational-basis measurement, for example, changes no outcome at all. On a circuit that does not scramble, the gap is wider: 0.71 exact against 0.60 predicted. The tables are in the module documentation, and a regression test holds the depth-independence. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

dcharlot added 3 commits 2026-10-05 08:08:11 -04:00

wai-quantum 0.3.31: quantum-inspired optimisation, a re-runnable classical baseline ... 92f9e000cd

Optimisation is the claim made most often for quantum hardware, and each such
claim is a comparison with a classical solver. quantum_qio is that solver,
deterministic and re-checkable:

- Ising problems (sparse symmetric J, fields h) and Max-Cut (J = -w), a G-set
  reader, and the +-1 all-to-all spin glass.
- Simulated bifurcation from its published update rules: symplectic Euler,
  perfectly inelastic walls, ballistic (f = Jx) and discrete (f = J sgn x)
  forces, and both heated (arXiv:2203.08361), with c0 = c1 / (rms(J) sqrt N).
  Fields ride on an extra spin fixed at +1.
- Eight trials advance per pass over the couplings; each trial's arithmetic is
  unchanged, so a batched run leaves the same final positions bit for bit as
  one trial at a time. The test fails on a 1e-15 change to one lane.
- Simulated annealing (single-spin Metropolis, linear beta), single-flip
  descent, an exact Gray-code referee up to 34 spins (threaded), and
  step-to-solution S = N_s ln 0.01 / ln(1 - P).

Checked:
- every variant and annealing reach the exact ground state on 20-spin glasses;
- G-set best-known cuts reached (seed 1): G1 11624 (39% of 2,000-step
  trials), G2 11620, G3 11622, G11 564, G14 3064 (annealing; bifurcation
  stops at 3062), G22 13359, G43 6660; G55 reaches 10292 of 10299;
- the published dense-glass step (1.1) is unstable on sparse Max-Cut graphs:
  the stiffest mode has w^2 = a0(a0 - c0 mu_min) and mu_min = -degree, so G1's
  limit is near 0.88; documented, with the settings that work.

CLI: wai-quantum qio (--gset FILE | --sk N) --method dsb|bsb|hdsb|hbsb|sa|exact
[--descend --target --spins --json]. Example: qio_gset. Pinned bit for bit
natively and under wasm32-wasip2 (16 pins); CLI output byte-identical across
the two, threaded or not, G-set files included.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

wai-quantum 0.3.32: union-find decoding in near-linear time ... e76acbe9e5

A fault-tolerant machine must decode each syndrome round before the next
arrives. Exact matching (quantum_match) is the accuracy reference; quantum_uf
is the fast decoder over the same matching graph:

- Grow: every fired detector starts a cluster; odd clusters (not on the
  boundary) grow all frontier edges at once, on the graph's integer weights,
  an edge between two growing clusters at double speed.
- Merge: a filled edge joins its clusters; a cluster on the boundary stops.
- Peel: each cluster's filled edges as a spanning tree, rooted at the
  boundary when it holds it, peeled leaves-first into a correction.
- Growth is event-driven: each frontier edge carries a fill time, and a merge
  reschedules only the frontier of a side whose growth changed. Fill times
  never run backwards, so events sit in a radix heap at amortised O(1).
  Profiling put 45% of the time in a binary heap before; the radix heap
  halved the time per shot at d = 21. Time now grows linearly with fired
  detectors, about 1.1-1.5 us each from d = 9 to d = 21.
- A reusable workspace resets in time proportional to the last shot.

Checked:
- every single edge of d = 3 and d = 5 graphs is corrected exactly;
- the event-driven schedule agrees with a round-by-round one (kept in the
  tests) on 2,000 shots at each of three distances, up to tie order;
- on the same 20,000 shots, union-find fails 0.92-1.18x as often as exact
  matching (d = 3-9, p = 0.003-0.007), 77x faster at d = 11; its threshold sits
  a little lower (at p = 0.007 its failures stop falling with distance).

CLI: wai-quantum qec --decoder uf|exact. Example: uf_bench. Pinned bit for bit
natively and under wasm32-wasip2 (17 pins); CLI output byte-identical.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

wai-quantum 0.3.33: noisy circuits, and where the digital error model is off ... 0d0211af6e

Every quantum-advantage experiment reports a fidelity measured on noisy
hardware. For random-circuit sampling that is a cross-entropy (XEB) estimate,
read against a digital error model: the product of every gate's and
readout's success probability. quantum_noisy checks that arithmetic:

- Trajectories on the state vector: after each gate, one branch of each noise
  channel on the qubits it touched. Pauli channels (depolarizing, dephasing)
  stay unitary; amplitude damping jumps with probability gamma P(1) and is
  renormalised; readout flips each measured bit. Seeded per trajectory and
  reduced in order, so a run is the same bits at any thread count.
- The exact density matrix up to 12 qubits, held as a state on 2n qubits
  (U (x) U* for a gate, the Kraus sum for a channel), on the same kernels.
- linear_xeb (sampled, with a standard error), exact_xeb (the infinite-sample
  value), ideal_xeb (the circuit's own, 1 when fully scrambling) and
  digital_fidelity.

Checked:
- 4,000 trajectories land within four standard errors of the exact
  expectations, every channel on; sampled bitstrings follow the exact
  distribution with readout flips; sampled XEB converges to the exact value.
- The digital model understates the exact (normalised) XEB of 8-qubit
  brickwork circuits by 3.5-4.5% at two-qubit error 0.005 and 7-10% at 0.01,
  and the margin does not grow over 4 to 48 layers: a fixed set of errors
  near the measurement that the model counts as fatal, not a per-gate bias.
  On a circuit that does not scramble the gap is wider (0.71 vs 0.60).

CLI: wai-quantum noisy --file CIRCUIT.jsonl [noise] [--expect] [--xeb]
[--exact] [--json]. Pinned bit for bit natively and under wasm32-wasip2 (18
pins); CLI output byte-identical, threaded or not.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

dcharlot added 2 commits 2026-10-05 09:08:44 -04:00

wai-quantum 0.3.32: close the optimisation pin's test function ... 79dd50d20b

The restack onto 0.3.31 left tests/bit_identity.rs with an unclosed function,
so the integration tests did not compile (the linux energy-receipts job). The
fix was staged after the commit by mistake. Also trims the union-find vs exact
matching comparison to 6,000 shots per distance (about 25 s in debug).

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Merge branch 'wai-quantum-uf' into wai-quantum-noisy 7211402746

dcharlot referenced this pull request 2026-10-05 09:50:45 -04:00

wai-quantum 0.3.34: ground states of spin chains by DMRG, held to exact answers #159

dcharlot merged commit 0fbc39d90e into main 2026-10-05 11:03:17 -04:00

dcharlot referenced this pull request from a commit 2026-10-05 11:03:18 -04:00

Merge pull request 'wai-quantum 0.3.33: noisy circuits, and where the digital error model is off' (#156) from wai-quantum-noisy into main

dcharlot referenced this pull request 2026-10-05 11:14:04 -04:00

wai-quantum 0.3.35: time evolution of spin chains by TDVP, held to exact answers #160

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

Transaction-Science/open-standards!156

No description provided.