Clarify CDR training docs regarding the use of a markov chain monte carlo #2130
Conversation
natestemen
added
documentation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #2130 +/- ##
=======================================
Coverage 98.19% 98.19%
=======================================
Files 88 88
Lines 4166 4166
=======================================
Hits 4091 4091
Misses 75 75 ☔ View full report in Codecov by Sentry. |
There was a problem hiding this comment.
Nice work @natestemen!
A few minor suggestions, otherwise LGTM!
docs/source/guide/cdr-4-low-level.md
Outdated
| Similarly to ZNE and PEC, CDR is divided in two main stages: first, one of circuit generation and a second for inference of the mitigated value. | ||
| In CDR, the generation of quantum circuits is different, as it involves the generation of training circuits. | ||
|
|
||
| ```{warning} |
There was a problem hiding this comment.
Not a big deal, but I would make this a note instead of a warning.
| ```{warning} | |
| ```{note} |
There was a problem hiding this comment.
I thought warning since it's likely an unexpected difference (hence the issue), but I think a note is fine too. Changed.
|
|
||
| The CDR workflow Figure above shows a schema of the implementation of CDR in Mitiq. Similarly to ZNE and PEC, also CDR in Mitiq is divided in two main stages: The first one of circuit generation and the second for inference of the mitigated value. However, in CDR, the generation of quantum circuits is different, as it involves the generation of training circuits. The division of CDR into training, learning and prediction stages is shown in the figure below. | ||
| The authors of {cite}`Czarnik_2021_Quantum` derive results with the use of the MCMC method, whereas Mitiq uses simpler approach presented in point 1. |
There was a problem hiding this comment.
Is there an indication of the effect on performance of using the simpler method instead of MCMC? If so, we should mention it.
There was a problem hiding this comment.
To my knowledge there has been no studies of how the circuit generation procedure effects the performance of CDR. Would be nice though!
co-authored-by: Misty Wahl <misty@unitary.fund>
Description
There is an important omission from our documentation regarding how Mitiq creates training circuits for Clifford Data Regression. There are two methods put forth in the original paper, yet the authors choice is different from the one in Mitiq. This PR adds clarification for that detail.
fixes #1393