[Feature] Additional features for error mitigation protocols #56

sungwoo-pasqal · 2025-02-24T14:46:32Z

Closes #30

Adds exponential decay extrapolation in analog ZNE method
Adds hamming distance that increases the sparsity in the MTHREE method
Adds finite random sampling in the Twirl method

RolandMacDoland

Thanks a lot @sungwoo-pasqal, it feels much cleaner :) Some comments from me but generally I think it'd be good to integrate the subspace redistribution logic into the function that creates the confusion matrix. Also maybe best to use hamming whenever possible. In other parts of Qadence, ham is a shortcut for Hamiltonian. This might cause confusion.

docs/error_mitigation/incoherent_error_mitigation.md

RolandMacDoland · 2025-02-27T16:04:31Z

docs/error_mitigation/incoherent_error_mitigation.md

@@ -29,13 +31,13 @@ model = QuantumModel(
    circuit=circuit, observable=observable, backend=BackendName.PULSER, diff_mode=DiffMode.GPSR
 )

-for data_points in [2,5]:
-    options = {"stretches": torch.linspace(1, 3, data_points)}
+for data_points, zne_type in zip([2,5], ["poly", "exp"]):


Does this presuppose the number of datapoints for each ZNE method ?

Yes. The exponential method needs at least three data points, whereas poly needs two. I'll put that stated in the docs.

RolandMacDoland · 2025-02-27T16:12:57Z

docs/error_mitigation/readout_mitigation.md

+confusion_matrix_subspace = normalized_subspace_kron(noise_matrices, observed_prob.nonzero()[0])
+
+# we consider a small hamming distance for this method and set it to 3
+confusion_matrix_subspace_ham = ham_dist_redistribution(


Instead of another function, would it make sense to integrate it into normalized_subspace_kron ? If I understand correctly, you use this latter function to generate a confusion matrix. But maybe if there is an extra ham_dist parameter that could trigger the subspace redistribution without having to rely on some logic defined elsewhere ?

RolandMacDoland · 2025-02-27T16:15:14Z

docs/error_mitigation/readout_mitigation.md

+confusion_matrix_subspace_ham = ham_dist_redistribution(
+                    confusion_matrix_subspace, ham_dist=3
+                )
+p_corr_mthree_gmres_ham = gmres(confusion_matrix_subspace_ham, input_csr.toarray())[0]


hm these variables names are a bit hard to understand. Could you please use more meaningful names or comment what you're doing here ?

RolandMacDoland · 2025-02-27T16:18:49Z

qadence_protocols/mitigations/analog_zne.py

@@ -27,6 +30,42 @@ def zne(noise_levels: Tensor, zne_datasets: list[list]) -> Tensor:
    return torch.tensor(poly_fits)


+def exp_func(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:


Suggested change

def exp_func(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:

def exp_fn(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:

Nothing major but there are few places in Qadence where we use fn for functions.

RolandMacDoland · 2025-02-27T16:20:41Z

qadence_protocols/mitigations/analog_zne.py

+    elif zne_type == "poly" or zne_type is None:
+        zne_func = zne
+    else:
+        raise ValueError("analog zne supports only polynomial or exponential extrapolation.")


Suggested change

raise ValueError("analog zne supports only polynomial or exponential extrapolation.")

raise ValueError(f"Analog ZNE supports only polynomial or exponential extrapolation. Got {zne_type}.")

RolandMacDoland · 2025-02-27T16:21:58Z

qadence_protocols/mitigations/readout.py

@@ -52,6 +52,34 @@ def normalized_subspace_kron(noise_matrices: npt.NDArrary, subspace: npt.NDArray
    return conf_matrix


+def ham_dist_redistribution(conf_matrix: npt.NDArray, ham_dist: int) -> npt.NDArray:


If you use npt.NDArray here, I'd suggest to keep it consistent elsewhere too.

RolandMacDoland · 2025-02-27T16:23:22Z

qadence_protocols/mitigations/readout.py

@@ -237,6 +265,15 @@ def mitigation_minimization(

        elif optimization_type == ReadOutOptimization.MTHREE:
            confusion_matrix_subspace = normalized_subspace_kron(noise_matrices, p_raw.nonzero()[0])
+
+            ham_dist = options.get("ham_dist")


Suggested change

ham_dist = options.get("ham_dist")

ham_dist = options.get("ham_dist", None)

RolandMacDoland · 2025-02-27T16:23:33Z

qadence_protocols/mitigations/readout.py

@@ -237,6 +265,15 @@ def mitigation_minimization(

        elif optimization_type == ReadOutOptimization.MTHREE:
            confusion_matrix_subspace = normalized_subspace_kron(noise_matrices, p_raw.nonzero()[0])
+
+            ham_dist = options.get("ham_dist")
+            if ham_dist:


Suggested change

if ham_dist:

if ham_dist is not None:

rajaiitp · 2025-02-27T18:09:43Z

qadence_protocols/mitigations/analog_zne.py

+
+        except RuntimeError:
+            print("Warning: Optimal parameters not found, using last known value.")
+            exp_fits.append(dataset_np[-1])


why do we choose to append the last value, and not something like Nan

rajaiitp · 2025-02-27T18:12:06Z

qadence_protocols/mitigations/analog_zne.py

+    if zne_type == "exp":
+        zne_func = zne_exp
+    elif zne_type == "poly" or zne_type is None:
+        zne_func = zne


If this is doing a poly fit, can we call this zne_poly rather than just zne

rajaiitp · 2025-02-27T18:25:08Z

qadence_protocols/mitigations/readout.py

+
+        if partial_sum > 0:  # Avoid division by zero
+            for row in valid_rows:
+                redistributed_matrix[row, col] = conf_matrix[row, col] / partial_sum


isnt it possible to just get the sub_matrix
redistributed_matrix = conf_matrix[valid_rows, col] or something ?

this one is m3 with increased hamming distance subspace confusion matrix, correct ? I was under the impression it would be a square matrix with both rows and columns matching and subsampled

I dont think partial sum can add to 0.. entries of confusion matrix are positive

rajaiitp · 2025-02-27T18:35:09Z

tests/test_readout.py

+
+    confusion_matrix_subspace = normalized_subspace_kron(noise_matrices, observed_prob.nonzero()[0])
+
+    confusion_matrix_subspace_ham = ham_dist_redistribution(confusion_matrix_subspace, 3)


i think you should first get the rows surrounding the non zero locations, expand by hamming distance and then call normalized_subspace over these locations

rajaiitp · 2025-02-27T18:42:17Z

qadence_protocols/mitigations/twirl.py

+    # Random sample gate strings for scalability
+    twirl_samples = options.get("twirl_samples")
+    if isinstance(twirl_samples, int) and twirl_samples > 0:
+        twirls = random.sample(twirls, twirl_samples)


It would be clever to not sample from the whole twirl space, but rather generate as much needed, as that would be scalable. Not sure if this is straightforward here, feel free to ignore

added exponential analog zne

4c39d33

sungwoo-pasqal added the feature label Feb 24, 2025

sungwoo-pasqal self-assigned this Feb 24, 2025

sungwoo-pasqal changed the title ~~[Feature] Additional features for readout mitigation protocols~~ [Feature] Additional features for error mitigation protocols Feb 25, 2025

sungwoo-pasqal added 2 commits February 27, 2025 06:13

hamming distance mthree added

b0dfcf0

add twirl finite sampling

7507b5d

sungwoo-pasqal marked this pull request as ready for review February 27, 2025 15:36

sungwoo-pasqal requested review from rajaiitp and RolandMacDoland February 27, 2025 15:36

RolandMacDoland reviewed Feb 27, 2025

View reviewed changes

rajaiitp requested changes Feb 27, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Feature] Additional features for error mitigation protocols #56

[Feature] Additional features for error mitigation protocols #56

sungwoo-pasqal commented Feb 24, 2025 •

edited

Loading

RolandMacDoland left a comment

RolandMacDoland Feb 27, 2025

sungwoo-pasqal Feb 28, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

RolandMacDoland Feb 27, 2025

rajaiitp Feb 27, 2025

rajaiitp Feb 27, 2025

rajaiitp Feb 27, 2025

rajaiitp Feb 27, 2025

rajaiitp Feb 27, 2025

		@@ -27,6 +30,42 @@ def zne(noise_levels: Tensor, zne_datasets: list[list]) -> Tensor:
		return torch.tensor(poly_fits)


		def exp_func(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:

	def exp_func(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:
	def exp_fn(x: np.ndarray, a: float, b: float, c: float) -> np.ndarray:

	raise ValueError("analog zne supports only polynomial or exponential extrapolation.")
	raise ValueError(f"Analog ZNE supports only polynomial or exponential extrapolation. Got {zne_type}.")

		@@ -52,6 +52,34 @@ def normalized_subspace_kron(noise_matrices: npt.NDArrary, subspace: npt.NDArray
		return conf_matrix


		def ham_dist_redistribution(conf_matrix: npt.NDArray, ham_dist: int) -> npt.NDArray:

	ham_dist = options.get("ham_dist")
	ham_dist = options.get("ham_dist", None)


		confusion_matrix_subspace = normalized_subspace_kron(noise_matrices, observed_prob.nonzero()[0])

		confusion_matrix_subspace_ham = ham_dist_redistribution(confusion_matrix_subspace, 3)

[Feature] Additional features for error mitigation protocols #56

Are you sure you want to change the base?

[Feature] Additional features for error mitigation protocols #56

Conversation

sungwoo-pasqal commented Feb 24, 2025 • edited Loading

RolandMacDoland left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sungwoo-pasqal commented Feb 24, 2025 •

edited

Loading