feat: qudit noise model (2/5) (#1854) by bramathon · Pull Request #1859 · rigetti/pyquil

bramathon · 2026-07-04T12:40:29Z

Part of the effort to land the experimental noise model and simulation module (originally PR #1848, branch nonbreaking-noise-model) as a reviewable stack. This is PR 2 of 5, stacked on #1858 — review/merge in order. Closes #1854.

This PR introduces the new quax-backed qudit noise model under pyquil/noise/:

Channel, MeasurementChannel, ResetChannel, and CycleChannel (_channels.py), and the new NoiseModel container plus DepolarizingNoiseModel / CompositeNoiseModel and fidelity estimators (_noise_model.py).
The historical pyquil.noise public API is preserved: pyquil/noise.py is moved to pyquil/noise/_legacy_noise.py and re-exported from pyquil/noise/__init__.py, with the legacy NoiseModel/KrausModel now marked deprecated in favor of the quax-based model.

It also adds qudit DefGate support in quilbase.py (matrix dimensions may now be any perfect power, e.g. 3, 9, …, not just powers of two), which the qutrit noise channels and later simulators rely on, and enables float64 (jax_enable_x64) for the test suite.

A handful of ResetChannel tests verify behavior end-to-end through the density-matrix simulator; since that simulator lands in the next PR, those specific assertions are deferred to #1855 (they are restored there). Everything else in the noise module is fully tested here (test_noise_model.py).

Stack

Drop Python 3.9 #1852 — Drop Python 3.9
Qudit Noise model #1854 — Qudit noise model (this PR)
Experimental exact simulators #1855 — Experimental exact simulators
Trajectory simulator #1856 — Trajectory simulator
Dynamic shape trajectory simulator #1857 — Dynamic-shape trajectory simulator

Tracked in #1863.

github-actions · 2026-07-04T12:43:20Z

Bencher Report

Project	pyquil
Branch	1854-qudit-noise-model
Testbed	ci-runner-linux

⚠️ WARNING: No Threshold found!
Without a Threshold, no Alerts will ever be generated.
Latency (nanoseconds (ns))
Click here to create a new Threshold
For more information, see the Threshold documentation.
To only post results if a Threshold exists, set the --ci-only-thresholds flag.

Click to view all benchmark results

Benchmark	Latency	seconds (s)
test/benchmarks/test_program.py::test_copy_everything_except_instructions	📈 view plot ⚠️ NO THRESHOLD	9.66 s
test/benchmarks/test_program.py::test_instructions	📈 view plot ⚠️ NO THRESHOLD	2.54 s
test/benchmarks/test_program.py::test_iteration	📈 view plot ⚠️ NO THRESHOLD	2.55 s

🐰 View full continuous benchmarking report in Bencher

asaites

I've made several comments, which I hope will be useful. I might spend a little more time looking through _noise_model.py, but for the most part, I think I've gone through this pretty thoroughly (and am doing my best to keep up with the physics).

Most of my comments are more style than anything, but I have a few concerns. Overall, though, this looks pretty solid!

asaites · 2026-07-15T15:13:33Z

+        if rows < 2:
+            return 0
+        for base in range(2, rows + 1):
+            k = int(round(math.log(rows, base)))
+            if base**k == rows:
+                return k
+        return 1


This might not matter much for the expected values of rows/reasonable values of len(self.matrix), but it is very slow for large numbers. As a much faster alternative, you can calculate it by iteratively checking if $\sqrt[p]{r} \in \mathbb{Z}$ for $p \in \text{PRIME}$, stopping when $\sqrt[p]{r} < 2$. If so, it's a factor of k, and you can reduce the problem to finding the solution for that new base, then multiply its factors by the one you've found.

Concretely:

def _int_n_root(x: int, n: int) -> int | None: """Return the largest integer less than or equal to n-th root of x.""" # This helper quickly checks if `x` is a perfect `n-th` root, # and will do so without the numerical instability # of the floating point check. # There are faster algorithms for very large numbers, # but this is already O(log(x)), and simple to implement. low, high = 1, x while low <= high: mid = (low + high) // 2 if (mid_pow_n := mid ** n) == x: return mid elif mid_pow_n < x: low = mid + 1 else: high = mid - 1 return low - 1 SMALL_PRIMES = [2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61] def find_k(r: int) -> tuple[int, int]: """For positive integer `r`, return integers `(b, k)` such that `r = b**k` and `k` is the largest integer for any integer `b >= 1` that satisfies this relation. If `r` is not a perfect power, this function returns `(r, 1)`. """ if r < 2: return r, 1 # Check if `r` is a perfect p-th power of a small prime `p`. # When it is, reduce `r` to that p-th root and build up the exponent `k`. base, k = r, 1 primes = iter(SMALL_PRIMES) p = next(primes, None) while p: if (b := _int_n_root(base, p)) < 2: return base, k elif b ** p != base: p = next(primes, None) else: k *= p base = b raise ValueError("r is too large to efficiently find k")

This approach is $O(\log^{2}{n})$ versus $O(n^2)$.

On further thought, this entire concept has a huge problem - it doesn't support quarts. Besides qubits and qutrits this the most likely qudit size. I think I'll need to add a option to specify the qudit dims

Introduce the new noise model under pyquil/noise/ (Channel, MeasurementChannel, ResetChannel, CycleChannel and the quax-based NoiseModel), preserving the legacy pyquil.noise API via _legacy_noise.py. Add qudit DefGate support in quilbase and enable float64 for tests. Part of splitting PR #1848 into a reviewable stack. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

github-actions · 2026-07-17T09:57:29Z

☂️ Python Coverage

current status: ❌

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
8174	7101	87%	87%	🔴

New Files

File	Coverage	Status
pyquil/noise/init.py	100%	🟢
pyquil/noise/_channels.py	78%	🔴
pyquil/noise/_noise_model.py	76%	🔴
TOTAL	85%	🔴

Modified Files

File	Coverage	Status
pyquil/quilbase.py	94%	🟢
TOTAL	94%	🟢

updated for commit: 65aa919 by action🐍

asaites · 2026-07-20T14:13:27Z

+        For a matrix of dimension d^k, returns k where d is the smallest
+        integer base >= 2 such that rows = d^k.
+        """
+        decomposition = _integer_base_and_exponent(len(self.matrix))


This doesn't work if the exponent is composite.

asaites · 2026-07-20T14:13:56Z

+        """
+        decomposition = _integer_base_and_exponent(len(self.matrix))
+        if decomposition is None:
+            return 0


Shouldn't this be 1?

I believe this should only happen if n < 2 which means 0 qubits. But on further thought, this should probably just be an error.

asaites · 2026-07-20T14:43:47Z

+        return None
+    # Find the smallest prime factor; it is the qudit dimension (base).
+    base = n
+    for p in range(2, int(math.isqrt(n)) + 1):


The whole point of isqrt is that it returns an integer, so there's no need to wrap it in int.

This method requires less work than the previous approach, but still does significantly more work than the one I proposed. Again, I'm not sure how large realistic values of n are, so maybe this doesn't matter much, my previous suggestion was based on that context, and this doesn't much address that concern.

asaites · 2026-07-20T14:51:30Z

-    num_qubits = int(round(np.log2(hilbert_dim)))
+    hilbert_dim = round(float(np.sqrt(shape[0]))) if superoperator else int(shape[0])
+    num_qubits = round(float(np.log2(hilbert_dim)))
    if 2**num_qubits != hilbert_dim:


Why have you added float conversions?

Thought it was a type error but maybe not.

bramathon · 2026-07-20T20:09:13Z

I've done a substantial refactor to add support for Lindbladian channels. The way this works is as follows:

ChannelProtocol defines the protocol for a channel which provides a superoperator associated with a gate. There are two concrete classes for this: Channel and SuperopChannel. The Channel has the _LindbladianBacked mixin, which means that it not only has a superoperator, but a lindbladian which generates it. This is now our canonical Channel while SuperopChannel is relegated to special use cases where we don't know the Lindbladian.

ResetChannel receives the same treatment, becoming _LindbladianBacked and with two concrete implementations.

CycleChannel and MeasureChannel are unchanged.

This was referenced Jul 4, 2026

feat: experimental exact simulators (3/5) (#1855) #1860

Draft

Experimental noise model & simulation module (tracking) #1863

Open

asaites self-assigned this Jul 13, 2026

bramathon force-pushed the 1852-drop-python-39 branch from f1884bc to 414f175 Compare July 15, 2026 18:08

Base automatically changed from 1852-drop-python-39 to master July 16, 2026 19:06

asaites reviewed Jul 16, 2026

View reviewed changes

bramathon and others added 2 commits July 17, 2026 08:46

Address review comments

65aa919

bramathon force-pushed the 1854-qudit-noise-model branch from a2e40e8 to 65aa919 Compare July 17, 2026 09:52

Add Lindbladian channels

948937d

asaites reviewed Jul 20, 2026

View reviewed changes

Renames

f7fdf61

Uh oh!

Conversation

bramathon commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Stack

Uh oh!

github-actions Bot commented Jul 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

⚠️ WARNING: No Threshold found!

Uh oh!

asaites left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

github-actions Bot commented Jul 17, 2026

☂️ Python Coverage

Overall Coverage

New Files

Modified Files

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bramathon commented Jul 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bramathon commented Jul 4, 2026 •

edited

Loading

github-actions Bot commented Jul 4, 2026 •

edited

Loading