Noise models

[sophiaponte]

175

channel-qvm.lisp is an implementation of a qvm that requires the user to explicitly define a noise model. The noise model is a set of noise rules. The noise rules describe where different types of noise should be injected in a program. Each rule is made up of a noise-predicate (describes the 'where' part) and operation-elements (kraus operators describing the noise). During each transition, the channel-qvm checks to see if the current instruction matches any of the noise-rules. If there is a match, the channel-qvm applies the noise-rule's operation elems either before or after the instruction is applied.

(defun build-noise-model ()
  (let* ((pred1 (make-noise-predicate (match-strict-qubits (list 0 1)) 1 "after"))
          (pred2 (make-noise-predicate (match-strict-gates (list "X" "Z")) 1 "after"))
          (kraus-elems1 (list (generate-damping-kraus-ops 2 1)))
          (kraus-elems2 (list (generate-damping-dephasing-kraus-ops 3 1)))
          (rule1 (make-noise-rule pred1 kraus-elems1))
          (rule2 (make-noise-rule pred2 kraus-elems2)))
    (make-noise-model (list rule1 rule2))))

approx-qvm.lisp Is an implementation of a qubit level noise model qvm. The noise model for this qvm is specified by adding T1, T2 and fReadout values for each qubit. The approx-qvm will generate the kraus operators for these values (or assignment probabilities in the case of fRo) and save them. Then, after each gate application instruction, this qvm searches for kraus operators belonging to the current qubit for either T1 or T2 and will apply them.

(defun make-noisy-approx-qvm ()
  (let* ((qvm (make-instance 'approx-qvm :number-of-qubits 2 :avg-gate-time 1)))
    (set-qubit-t1 qvm 0 5)
    (set-qubit-t2 qvm 0 4)
    (set-qubit-t1 qvm 1 2)
    (set-qubit-t2 qvm 1 3)))

[braised-babbage]

Overall this is good progress. Most of the things I mentioned are just minor stylistic tweaks to make this look more "idiomatic". The only more substantial issue is the predicate stuff. Here I think predicates should be basically any boolean function of a quil:instruction. This will take a bit more writing to get working at once, but is more powerful and more sustainable than the previous pyquil approach.

[ecpeterson]

I gave this a once-over, and it looks like it's headed very much in the right direction. Please ping me to check back again after you've knocked out Erik's and Appleby's comments.

Exciting changes ahead for noise in the QVM!

[sophiaponte]

Recent updates:

• I put the NOISE-MODEL, NOISE-PREDICATE, and NOISE-RULE classes and related functions into a separate file (noise-models.lisp). I also moved related tests into tests/noise-model-tests.lisp

• I added DEPOLARIZING-NOISE to APPROX-QVM, as it is a frequently used to model QPU noise.

• I added another "built-in" predicate function (match-instr-idxs parsed-program &rest idxs). This returns true if the index of the current instr within parsed-program is a member of idxs.

EDIT:

• I also added a NAME slot to NOISE-PREDICATE and a function to print a noisy program with the predicate NAMEs printed in the correct places

[sophiaponte]

In this last review I mostly made changes to documentation and a few smaller stylistic things. However, I did have to make a pretty big change in BASIC-NOISE-QVM to more accurately generate the Kraus operators for damping and dephasing noise. I'll summarize the changes here:

• Damping and depolarizing noise is time dependent. Therefore, it doesn't make sense to precompute the kraus operators for this noise before running a qvm program. Now, the operators for these noise sources are calculated at every transition, and the kraus operators are parametrized by the elapsed time since the start of the program.
• I added an elasped-time slot to the basic-noise-qvm which is incremented by average-gate-time after every instruction, to address the above point.
• Dephasing noise is calculated from t-phi. I added a slot to the qvm so that users can provide t-phi values for each qubit. If these values are not provided, then t-phi is calculated from T1 and T2, if these are available.

[sophiaponte]

I went through the src files again to catch any formatting issues, and I think I've cleaned everything up at this point. The last thing that I need to do is squash my commits, and then I think this should be good to go unless anyone thinks otherwise :)