Fill diagonal elements of error_matrix from gradient estimators #85
Conversation
r-imai-quantum
force-pushed
the
error-matirx-gradient-estimator
branch
2 times, most recently
from
74237b9 to
2b1f55f
Compare
| var = ( | ||
| (estimates[2 * i].error ** 2) + (estimates[2 * i + 1].error ** 2) | ||
| ) / delta | ||
| err_diag.append(np.sqrt(var)) |
There was a problem hiding this comment.
This seems to be
There was a problem hiding this comment.
Thank you for pointing this out!
I misremembered the propagation rule of error.
The latter one is correct.
| @@ -62,7 +64,7 @@ def numerical_gradient_estimates( | |||
| Returns: | |||
| The estimated values (can be accessed with :attr:`.values`) with errors | |||
| of estimation (can be accessed with :attr:`.error_matrix`). Currently, | |||
| :attr:`.error_matrix` returns `None`. | |||
| non-diagonal elements of :attr:`.error_matrix` are filled with zeros. | |||
There was a problem hiding this comment.
Saying "are filled with zeros" sounds to me like "we don't know the exact values but we put zeros at the moment", but I guess the covariances are exactly zero in this case since each element is estimated indepedently.
There was a problem hiding this comment.
I was a bit confused but, as you guessed, the covariance are exactly zero.
So I've removed the sentence.
| for p, c in params_and_coefs: | ||
| g += estimates_dict[p].value * c | ||
| var += (estimates_dict[p].error ** 2) * abs(c) |
There was a problem hiding this comment.
I guess
| var += (estimates_dict[p].error ** 2) * abs(c) | |
| var += (estimates_dict[p].error ** 2) * abs(c) ** 2 |
r-imai-quantum
force-pushed
the
error-matirx-gradient-estimator
branch
from
2b1f55f to
e4d3868
Compare
|
@kwkbtr |
In the current implementation, gradient estimators do not return error_matrix.
This PR contains changes that make estimators return a matrix where the diagonal elements are filled with its error.
Note that non-diagonal elements are filled with zeros without calculation.