Missing shots break the expectation value computation

[mrossinek]

Environment

• qiskit-addon-utils version: 0.3.0
• Python version: Any
• Operating system: Any

What is happening and why is it wrong?

The edge case handling in which all shots have been filtered out is broken. More specifically, this code is causing issues:

qiskit-addon-utils/qiskit_addon_utils/exp_vals/expectation_values.py

Lines 398 to 401 in 9b92faf

	# Edge case of counts dict containing outcomes but with total shots, eg {"0": 0}.
	no_shots = denom == 0
	expvals_each_term[~no_shots] /= denom[..., np.newaxis]

How can we reproduce the issue?

Try to compute an expectation value for a set of data in which the post-selection removes all shots for any observable term.

I will work on putting together a simple to reproduce example.

Traceback

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[217], line 1
----> 1 res = post_process_conv(exec_result[0], steps=steps, gamma=gamma, ps="node", trex=True)

Cell In[212], line 27, in post_process_conv(datum, steps, gamma, ps, trex)
     24 results = []
     25 for i in range(steps, num_randomizations + 1, steps):
     26     # Compute mitigated expvals with post-selectoion
---> 27     res = executor_expectation_values(
     28         meas[:i],
     29         reverser,
     30         meas_basis_axis,
     31         avg_axis=avg_axis,
     32         measurement_flips=flips[:i],
     33         pauli_signs=signs[:i] if signs is not None else None,
     34         postselect_mask=mask[:i] if mask is not None else None,
     35         rescale_factors=trex_scale_factors if trex else None,
     36         gamma_factor=gamma,
     37     )
     38     results.append(res[0])
     39 return results

File lib64/python3.13/site-packages/qiskit_addon_utils/exp_vals/expectation_values.py:185, in executor_expectation_values(bool_array, basis_dict, meas_basis_axis, avg_axis, measurement_flips, pauli_signs, postselect_mask, gamma_factor, rescale_factors)
    180 basis_rescale_factors = (
    181     rescale_factors[meas_basis_idx] if rescale_factors is not None else None
    182 )
    184 ## AVERAGE OVER SHOTS:
--> 185 (means, standard_errs) = _bitarray_expectation_value(
    186     barray_this_basis,
    187     observables,
    188     shots=num_kept,
    189     rescale_each_observable=basis_rescale_factors,
    190 )
    192 variances = standard_errs**2
    193 del standard_errs

File lib64/python3.13/site-packages/qiskit_addon_utils/exp_vals/expectation_values.py:401, in _bitarray_expectation_value(outcomes, observables, shots, rescale_each_observable)
    398 # Edge case of counts dict containing outcomes but with total shots, eg {"0": 0}.
    399 no_shots = denom == 0
--> 401 expvals_each_term[~no_shots] /= denom[..., np.newaxis]
    402 sq_expvals_each_term[~no_shots] /= denom[..., np.newaxis]
    403 expvals_each_term[no_shots] = np.nan

ValueError: operands could not be broadcast together with shapes (14,1) (16,1) (14,1)

Any suggestions?

We need to update our handling of no_shots. While the intention was good, the implementation is not correct as it stands.

It could be as simple as also using no_shots for denom but I need to verify the axis handling.

If that works, we still need to account for the potential presence of np.nan later on in the code, since we end up summing expvals_each_term to compute the total observable expectation value. If we don't account for np.nan there, we will simply get out np.nan for the summed value which is not what we want.