WIP: crosstalk-free GST by nkoskelo · Pull Request #641 · sandialabs/pyGSTi

nkoskelo · 2025-08-19T17:37:52Z

TODO: Add description.

…ature-crosstalk-free-gst

…c parent class of CrosstalkFreeExperimentDesign.

…imentdesign functions

…gates with implicit availability

… just raise an more informative error). Add a copy in PermutationOperator.pp_braiding_operators.

…VXOPT

…ys compute model violation w.r.t. logl objective function, rather than whatever might have been used as the final objective function.

…s. These are hacky changes. A proper change would be to use try-catch if encountering linalg errors.

…functions only use idles with explicit target qubits, rather than implicit global idles

… explicitmodel.py. Update convert_members_inplace so it can handle failures of conversion when dealing with embedded ops.

…test_mpi.py

… in matrixforwardsim.py.

pygsti/baseobjs/statespace.py

rileyjmurray · 2025-08-19T18:04:40Z

pygsti/circuits/circuit.py

    if isinstance(x, _Label): return x
    # # do this manually when desired, as it "boxes" a circuit being inserted
-    #elif isinstance(x,Circuit): return x.to_circuit_label()
+    elif isinstance(x,Circuit): return x.to_label()


Neither Nick nor I remember making this change. The commented-out branch was broken. The uncommenting seems valid based on the to_label member function of Circuit objects.

pygsti/report/factory.py

pygsti/tools/optools.py

…atic method

… to a static method" This reverts commit 8226079.

pygsti/layouts/matrixlayout.py

nkoskelo · 2025-08-19T20:47:14Z

pygsti/circuits/circuitlist.py

Tensor circuit construction should retain knowledge of original lane structure.

nkoskelo · 2025-08-19T21:24:02Z

pygsti/circuits/circuit.py

        #self._times = None  # for FUTURE expansion
        self.auxinfo = {}  # for FUTURE expansion / user metadata
+        self.saved_auxinfo = {}
+        self.saved_auxinfo["lanes"] = {tuple(line_labels): self._labels}


We are saving off the lane structure here so that it can be used in the future if the error models are cross-talk free.

The changes in this PR were originally made for crosstalk-free GST work (see #641). We've factored out those changes into this PR to reduce the size of #641 (and get these incidental changes in sooner rather than later). Changes are overwhelmingly type annotations. --------- Co-authored-by: Nick Koskelo <koskelo2@illinois.edu> Co-authored-by: nkoskelo <129830924+nkoskelo@users.noreply.github.com>

rileyjmurray

@nkoskelo now that we've got the xfgst-incidentals merged in we'll want to create a new (smaller scope) crosstalk-free GST PR. The new PR should include everything that's useful in XFGST modeling except the things related to a new forward simulator. I've left comments on the code that matches that description. Please do a pass of your own and let me know if you agree with what I've highlighted and/or if you think I missed something.

Once we agree on the plan here we can close this PR and create a new branch starting from develop!

rileyjmurray · 2026-01-29T14:46:57Z

pygsti/models/implicitmodel.py

+    def create_explicit(self, composedops_as_views=False, spam_type='full TP', op_type='CPTPLND'):
+        from pygsti.models import ExplicitOpModel
+        if isinstance(composedops_as_views, str):
+            # Old positional argument behavior.
+            spam_type = composedops_as_views
+            op_type = composedops_as_views
+            composedops_as_views = False
+
+        state = self.__getstate__()
+        state['povms'] = state['povm_blks']['layers']
+        state['preps'] = state['prep_blks']['layers']
+        opdict = _OrderedMemberDict(None, spam_type, None, dict())
+        opdict.update(state['_opcaches']['complete-layers'])
+        state['operations'] = opdict
+
+        for v in state['operations'].values():
+            v.unlink_parent(force=True)
+        for v in state['povms'].values():
+            v.unlink_parent(force=True)
+        for v in state['preps'].values():
+            v.unlink_parent(force=True)
+
+        eom = ExplicitOpModel(self.state_space)
+        eom.preps.update(state['preps'])
+        eom.povms.update(state['povms'])
+        eom.operations.update(state['operations'])
+
+        eom.convert_members_inplace(
+            to_type=spam_type, categories_to_convert=('preps', 'povms'),
+            allow_smaller_pp_basis=True, flatten_structure=True
+        )
+        eom.convert_members_inplace(
+            to_type=op_type, categories_to_convert=('operations',),
+            allow_smaller_pp_basis=True
+        )
+        if composedops_as_views:
+            from pygsti.modelmembers.operations import ComposedOp
+            allop_keys = eom.operations.keys()
+            for k in allop_keys:
+                curr_op = eom[k]
+                if isinstance(curr_op, ComposedOp) and all([subk in allop_keys for subk in k]):
+                    view_op = ComposedOp([eom[subk] for subk in k])
+                    eom[k] = view_op
+            eom._rebuild_paramvec()
+        return eom


Worth merging.

rileyjmurray · 2026-01-29T14:48:27Z

pygsti/models/explicitmodel.py

+    def convert_members_inplace(self, to_type,
+            categories_to_convert='all', labels_to_convert='all',
+            ideal_model=None, flatten_structure=False, set_default_gauge_group=False,
+            cptp_truncation_tol= 1e-7, spam_cp_penalty = 1e-7, allow_smaller_pp_basis=False
+        ):
        """
        TODO: docstring -- like set_all_parameterizations but doesn't set default gauge group by default
+
+        allow_smaller_pp_basis : bool
+
+            Setting allow_smaller_pp_basis=True allows dimension mismatches between
+            this ExplicitOpModel's operations and the dimensions we'd expect for
+            operations based on the properties of self.basis.
+
+            We can ONLY handle mismatches when self.basis.name indicates a Pauli product basis
+            or a tensor product thereof. We handle a failed conversion by trying a second time,
+            passing in the string literal `pp` instead of `self.basis` to _op.convert(...).
+
        """
        if isinstance(categories_to_convert, str): categories_to_convert = (categories_to_convert,)
+        assert all(c in ['all', 'ops', 'operations', 'instruments', 'preps', 'povms'] for c in categories_to_convert)
+        fallback_basis = '' if not allow_smaller_pp_basis else self.basis.name.replace('pp','').replace('*','') + 'pp'
+        ideal_model = ModelView.cast(ideal_model)
+        roster = Roster(labels_to_convert)
        if any([c in categories_to_convert for c in ('all', 'ops', 'operations')]):
-            for lbl, gate in self.operations.items():
-                if labels_to_convert == 'all' or lbl in labels_to_convert:
-                    ideal = ideal_model.operations.get(lbl, None) if (ideal_model is not None) else None
-                    self.operations[lbl] = _op.convert(gate, to_type, self.basis, ideal, flatten_structure, cptp_truncation_tol)
+            op_items = [(k,v) for (k,v) in self.operations.items() if k in roster]
+            for lbl, gate in op_items:
+                ideal = ideal_model.operations[lbl]
+                try:
+                    op = _op.convert(gate, to_type, self.basis, ideal, flatten_structure, cptp_truncation_tol)
+                except ValueError as e:
+                    if not fallback_basis == 'pp':
+                        raise e
+                    op = _op.convert(gate, to_type, 'pp', ideal, flatten_structure, cptp_truncation_tol)
+                self.operations[lbl] = op
        if any([c in categories_to_convert for c in ('all', 'instruments')]):
-            for lbl, inst in self.instruments.items():
-                if labels_to_convert == 'all' or lbl in labels_to_convert:
-                    ideal = ideal_model.instruments.get(lbl, None) if (ideal_model is not None) else None
-                    self.instruments[lbl] = _instrument.convert(inst, to_type, self.basis, ideal, flatten_structure)
+            inst_items = [(k,v) for (k,v) in self.instruments.items() if k in roster]
+            for lbl, inst in inst_items:
+                ideal = ideal_model.instruments[lbl]
+                self.instruments[lbl] = _instrument.convert(inst, to_type, self.basis, ideal, flatten_structure)
        if any([c in categories_to_convert for c in ('all', 'preps')]):
-            for lbl, prep in self.preps.items():
-                if labels_to_convert == 'all' or lbl in labels_to_convert:
-                    ideal = ideal_model.preps.get(lbl, None) if (ideal_model is not None) else None
-                    self.preps[lbl] = _state.convert(prep, to_type, self.basis, ideal, flatten_structure, cp_penalty=spam_cp_penalty)
+            prep_items = [(k,v) for (k,v) in self.preps.items() if k in roster]
+            for lbl, prep in prep_items:
+                ideal = ideal_model.preps[lbl]
+                self.preps[lbl] = _state.convert(prep, to_type, self.basis, ideal, flatten_structure, cp_penalty=spam_cp_penalty)
        if any([c in categories_to_convert for c in ('all', 'povms')]):
-            for lbl, povm in self.povms.items():
-                if labels_to_convert == 'all' or lbl in labels_to_convert:
-                    ideal = ideal_model.povms.get(lbl, None) if (ideal_model is not None) else None
-                    self.povms[lbl] = _povm.convert (povm, to_type, self.basis, ideal, flatten_structure, cp_penalty=spam_cp_penalty)
+            povm_items = [(k,v) for (k,v) in self.povms.items() if k in roster]
+            for lbl, povm in povm_items:
+                ideal = ideal_model.povms[lbl]
+                self.povms[lbl] = _povm.convert(povm, to_type, self.basis, ideal, flatten_structure, cp_penalty=spam_cp_penalty)


Maybe worth upstreaming. Would need the Roster and ModelView classes defined at the top of the file if so.

rileyjmurray · 2026-01-29T14:49:18Z

pygsti/circuits/circuitlist.py

+
+    def tensor_circuits(self, other_circuitlist, new_name=None):
+        assert len(self) == len(other_circuitlist)
+        circuits = []
+        for c1,c2 in zip(self._circuits, other_circuitlist._circuits):
+            circuits.append(c1.tensor_circuit(c2))
+        out = CircuitList(circuits, name=new_name)
+        return out


Maybe worth merging.

rileyjmurray · 2026-01-29T14:50:14Z

pygsti/circuits/split_circuits_into_lanes.py

TODO: determine if we want to keep these new functions.

Can see these functions in action in test/unit/objects/test_circuit_splitting.py.

rileyjmurray · 2026-01-29T14:52:11Z

pygsti/forwardsims/matrixforwardsim.py

@@ -739,6 +751,9 @@ def _compute_product_cache(self, layout_atom_tree, resource_alloc):
        eval_tree = layout_atom_tree
        cacheSize = len(eval_tree)
        prodCache = _np.zeros((cacheSize, dim, dim), 'd')
+        # ^ This assumes assignments prodCache[i] = <2d numpy array>.
+        #   It would be better for this to be a dict (mapping _most likely_
+        #   to ndarrays) if we don't need slicing or other axis indexing.
        scaleCache = _np.zeros(cacheSize, 'd')

        for iDest, iRight, iLeft in eval_tree:
@@ -752,7 +767,10 @@ def _compute_product_cache(self, layout_atom_tree, resource_alloc):
                else:
                    gate = self.model.circuit_layer_operator(opLabel, 'op').to_dense("minimal")
                    nG = max(_nla.norm(gate), 1.0)
+                    # ^ This indicates a need to compute norms of the operation matrices. Can't do this
+                    #   with scipy.linalg if gate is represented implicitly. 
                    prodCache[iDest] = gate / nG
+                    # ^ Indicates a need to overload division by scalars.
                    scaleCache[iDest] = _np.log(nG)
                continue

@@ -765,10 +783,12 @@ def _compute_product_cache(self, layout_atom_tree, resource_alloc):
            scaleCache[iDest] = scaleCache[iLeft] + scaleCache[iRight]

            if prodCache[iDest].max() < _PSMALL and prodCache[iDest].min() > -_PSMALL:
-                nL, nR = max(_nla.norm(L), _np.exp(-scaleCache[iLeft]),
-                             1e-300), max(_nla.norm(R), _np.exp(-scaleCache[iRight]), 1e-300)
+                nL = max(_nla.norm(L), _np.exp(-scaleCache[iLeft]), 1e-300)
+                nR = max(_nla.norm(R), _np.exp(-scaleCache[iRight]), 1e-300)
                sL, sR = L / nL, R / nR
-                prodCache[iDest] = _np.dot(sL, sR); scaleCache[iDest] += _np.log(nL) + _np.log(nR)
+                # ^ Again, shows the need to overload division by scalars.
+                prodCache[iDest] = sL @ sR
+                scaleCache[iDest] += _np.log(nL) + _np.log(nR)

        nanOrInfCacheIndices = (~_np.isfinite(prodCache)).nonzero()[0]  # may be duplicates (a list, not a set)
        # since all scaled gates start with norm <= 1, products should all have norm <= 1
@@ -839,6 +859,8 @@ def _compute_dproduct_cache(self, layout_atom_tree, prod_cache, scale_cache,

        tSerialStart = _time.time()
        dProdCache = _np.zeros((cacheSize,) + deriv_shape)
+        # ^ I think that deriv_shape will be a tuple of length > 2.
+        #   (Based on how swapaxes is used in the loop below ...)
        wrtIndices = _slct.indices(wrt_slice) if (wrt_slice is not None) else None

        for iDest, iRight, iLeft in eval_tree:
@@ -852,6 +874,9 @@ def _compute_dproduct_cache(self, layout_atom_tree, prod_cache, scale_cache,
                    #doperation = self.dproduct( (opLabel,) , wrt_filter=wrtIndices)
                    doperation = self._doperation(opLabel, wrt_filter=wrtIndices)
                    dProdCache[iDest] = doperation / _np.exp(scale_cache[iDest])
+                    # ^ Need a way to track tensor product structure in whatever's 
+                    #   being returned by self._doperation (presumably it's a tensor ...)
+
                continue

            tm = _time.time()
@@ -862,18 +887,31 @@ def _compute_dproduct_cache(self, layout_atom_tree, prod_cache, scale_cache,
            # since then matrixOf(circuit[i]) = matrixOf(circuit[iLeft]) * matrixOf(circuit[iRight])
            L, R = prod_cache[iLeft], prod_cache[iRight]
            dL, dR = dProdCache[iLeft], dProdCache[iRight]
-            dProdCache[iDest] = _np.dot(dL, R) + \
-                _np.swapaxes(_np.dot(L, dR), 0, 1)  # dot(dS, T) + dot(S, dT)
+            term1 = _np.dot(dL, R)
+            term2 = _np.swapaxes(_np.dot(L, dR), 0, 1) 
+            # ^ From the numpy docs on .dot :
+            #
+            #   If a is an N-D array and b is an M-D array (where M>=2),
+            #   it is a sum product over the last axis of a and the second-to-last axis of b:
+            #
+            #       dot(a, b)[i,j,k,m] = sum(a[i,j,:] * b[k,:,m])
+            #
+            dProdCache[iDest] = term1 +  term2  # dot(dS, T) + dot(S, dT)
+            # ^ We need addition of tensor-product-structured "doperators."
+
            profiler.add_time("compute_dproduct_cache: dots", tm)
            profiler.add_count("compute_dproduct_cache: dots")

            scale = scale_cache[iDest] - (scale_cache[iLeft] + scale_cache[iRight])
            if abs(scale) > 1e-8:  # _np.isclose(scale,0) is SLOW!
                dProdCache[iDest] /= _np.exp(scale)
                if dProdCache[iDest].max() < _DSMALL and dProdCache[iDest].min() > -_DSMALL:
+                    # ^ Need the tensor-product-structured "doperators" to have .max() and .min()
+                    #   methods.


These changes might be worth keeping.

rileyjmurray · 2026-01-29T14:53:53Z

pygsti/models/localnoisemodel.py

todo: figure out if we want to merge these changes

rileyjmurray · 2026-01-29T14:54:24Z

pygsti/protocols/crosstalkfreeexperimentdesign.py

we'll want to merge this

rileyjmurray · 2026-01-29T14:56:37Z

pygsti/tools/matrixtools.py

+    if len(inds) == 1:
        if add:
-            a[inds] += rhs  # all integers or slices behave nicely
+            a[inds] += rhs
        else:
-            a[inds] = rhs  # all integers or slices behave nicely
+            a[inds] = rhs
+    elif all([isinstance(i, (int, slice)) for i in inds]):
+        assert len(inds) == rhs.shape[1]
+        for ind, rhs_vec in zip(inds, rhs.T):
+            if add:
+                a[ind] += rhs_vec  # all integers or slices behave nicely
+            else:
+                a[ind]  = rhs_vec  # all integers or slices behave nicely


I think we'll want to keep this

rileyjmurray · 2026-01-29T14:57:27Z

test/unit/mpi/run_me_with_mpiexec.py

+pygsti.optimize.customsolve.CUSTOM_SOLVE_THRESHOLD = 10
 wcomm = MPI.COMM_WORLD
+print(f'Running with CUSTOM_SOLVE_THRESHOLD = {pygsti.optimize.customsolve.CUSTOM_SOLVE_THRESHOLD}')


This should be merged

rileyjmurray · 2026-01-29T15:00:42Z

test/unit/objects/test_forwardsim_on_implicitop_model.py

Probably want to keep this

rileyjmurray · 2026-03-13T20:35:07Z

@nkoskelo please see my most recent review. I'd like for you to create a new "Crosstalk-free GST infrastructure" PR based on the files indicated there. As soon as that PR is open we can/should close this one.

dhothem and others added 30 commits January 20, 2025 15:49

Adding in crosstalk-free GST eDesigns.

16e9309

Merge branch 'master' of https://github.com/sandialabs/pyGSTi into fe…

c483ebb

…ature-crosstalk-free-gst

style

56242ee

restore commented-out line with update to function name

d249c80

whitespace and docstring

dc964ad

add some utility functions to CircuitList class

8018e69

bugfixes in stich_circuits_by_germ_power_only. Remove HasProcessorSpe…

1d29ddb

…c parent class of CrosstalkFreeExperimentDesign.

undo changes to CircuitList. Add tqdm logging. Fix crosstalkfreeexper…

19fcf3c

…imentdesign functions

add PermutationOperator class. Remove outdated part of a docstring.

35e35fa

update ProcessorSpec.compute_2Q_connectivity() so it can handle idle …

3e3210f

…gates with implicit availability

undo imposition of certain connectivity structure for idle gates (now…

1253ee5

… just raise an more informative error). Add a copy in PermutationOperator.pp_braiding_operators.

ready for 3qubit xfgst

99392bf

replace assertion with warning

6b6b69e

switch to try MOSEK if available, falling back on Clarabel and then C…

c54c4d0

…VXOPT

add ability to skip named sections in construct_standard_report. Alwa…

a875d44

…ys compute model violation w.r.t. logl objective function, rather than whatever might have been used as the final objective function.

make computing some report quantities more robust to degenerate input…

eef4b44

…s. These are hacky changes. A proper change would be to use try-catch if encountering linalg errors.

make sure that circuits produced in crosstalk free experiment design …

689ffa7

…functions only use idles with explicit target qubits, rather than implicit global idles

style

1d66e1c

use separate 2q idle label

028e961

methods for (de)serialization of PermutationOperator

9d2890a

tweak meaning of force=True in ModelMember.unlink_parent

5644b20

leave a TODO

e96df9f

Clean up convert_members_inplace by adding two tiny helper classes to…

021fbf3

… explicitmodel.py. Update convert_members_inplace so it can handle failures of conversion when dealing with embedded ops.

add create_explicit method to LocalNoiseModel class

1821244

add option to change threshold at which customsolve is used. Updated …

35d0353

…test_mpi.py

comment where changes might be incorporated for faster forward sim

5a90e5c

logging

4f3ba4d

argument check

5f96a16

move LocalNoiseModel.create_explicit to ImplicitOpModel. Add comments…

dd86821

… in matrixforwardsim.py.

add features to create_explicit

7f0d9dd

merge develop

84b5356

rileyjmurray reviewed Aug 19, 2025

View reviewed changes

pygsti/baseobjs/statespace.py Show resolved Hide resolved

rileyjmurray reviewed Aug 19, 2025

View reviewed changes

Merge made memo set twice.

7992a6b

rileyjmurray reviewed Aug 19, 2025

View reviewed changes

pygsti/report/factory.py Outdated Show resolved Hide resolved

rileyjmurray reviewed Aug 19, 2025

View reviewed changes

pygsti/tools/optools.py Outdated Show resolved Hide resolved

rileyjmurray mentioned this pull request Aug 19, 2025

Upstream xfgst incidentals #643

Merged

rileyjmurray and others added 7 commits August 19, 2025 12:41

finish actually changing Circuit.from_cirq from a classmethod to a st…

8226079

…atic method

Revert "finish actually changing Circuit.from_cirq from a classmethod…

5a29f4b

… to a static method" This reverts commit 8226079.

Move daydic kron op to tools.

d4e5db6

Remove permutation op.

b3a226c

Remove unused_xfgst.py

72303c3

Reactivate test_errgenproptools.py

19d43d5

Remove unused code and commented out code.

525373e

nkoskelo commented Aug 19, 2025

View reviewed changes

pygsti/layouts/matrixlayout.py Show resolved Hide resolved

nkoskelo added 2 commits August 19, 2025 13:44

Revert changes to mapforward_calc_generic

bb78983

No actually though.

d0152d1

nkoskelo commented Aug 19, 2025

View reviewed changes

pygsti/circuits/circuitlist.py

Copy link

Contributor Author

nkoskelo Aug 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Tensor circuit construction should retain knowledge of original lane structure.

nkoskelo added 3 commits August 19, 2025 13:48

Matrix layout update.

58d97e0

Types for sequencetools

30ee59f

Add licensing info.

dd402cf

nkoskelo commented Aug 19, 2025

View reviewed changes

nkoskelo added 5 commits August 19, 2025 14:28

Remove code not needed for the PR.

3431730

Extract eval tree to new file.

b4f7044

Finish the split.

413ac67

Add a description for EvalTreeBasedUponLongestCommonSubstring

3bd478e

CollectionLCS Eval Tree documentation.

8f57778

rileyjmurray reviewed Jan 29, 2026

View reviewed changes

Conversation

nkoskelo commented Aug 19, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rileyjmurray left a comment

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!

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!

rileyjmurray commented Mar 13, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

nkoskelo commented Aug 19, 2025 •

edited

Loading