Hacker News Reader

Comment by andrewla | Hacker News Reader

Also of note is Gil Kalai's updated take on the experiment [1]. The heart of the complaint is that the quantum computer's solution to this problem relies on a calibration process, which is done on a classical computer and requires resources orders of magnitudes higher than the quantum portion of the computation:

> The Google experiment actually showed that a quantum computer running for 100 seconds PLUS a classic computer that runs 1000 hours can compute something that requires a classic computer 10 hours.

EDIT: Just as a note, lest you dismiss Kalai as an outsider/crackpot, you can look at the researchers who show up in the comments on the blog post.

[1] https://gilkalai.wordpress.com/2019/10/13/the-story-of-poinc...

View on HN

Crucial question - is the calibration a one time effort where the same calibration can be used to calculate any number of results within the calibrated space? In other words, can you amortize the calibration cost across 1,000 runs and come out an order of magnitude ahead?

See the comments on Kalai's post from Peter Shor, where Kalai attempts to address this, and my reply to mantap on this thread where I try to give my poor understanding of the critique.

I'll just say the line between being unfairly critical and raising valid points seem to be a very thin one.

Especially when Peter Shore is disagreeing with the critique (and I agree with him).

Saying that a quantum computation does not qualify because it included (a short) classic computation does not seem a fundamentally fair criticism to me.

"John Martinis explicitly confirmed for me that once the qubits are calibrated, you can run any circuit on them that you want."

https://www.scottaaronson.com/blog/?p=4372

Just saw this! As Aaronson mentions, coupled with the IBM work, if it does work out, it should be possible to verify that the 53-bit samples are actually valid through the cross entropy test, which would be very exciting.

If I understand his point correctly, it's not the calibration per se that is problematic, but whether the calibration means that it only performs for a small subspace of inputs, which raises the question how large is that subspace?

My understanding is that the calibration could only be feasibly carried out for smaller circuits where simulating the circuit classically was possible. At issue is whether the calibration is generalizable; whether the calibration carried out for a particular small circuit C was reused to simulate another circuit C', or whether fresh calibrations were carried out.

Page 17 of the supplemental data [1] shows the calibration process in schematic, and appears to indicate that the calibration was necessary for each input circuit C, since the graph showing the fidelity improving with iterations of the calibration is labelled "b, Data from a two-qubit XEB experiment." The water is admittedly muddy, and hopefully clarification will be forthcoming from some quarter to determine whether Kalai's particular critique has substance.

[1] https://static-content.springer.com/esm/art%3A10.1038%2Fs415...

Even if the calibration process "cheats", why should we care? It's not like being practical is the goal. After the calibration is done, you still have a device that can do something believed intractable on classical computers. If it is possible for a classical computer to get comparable results just from a long calibration process, then that means the cross entropy test is flawed/limited, which is a separate argument.

If you're going to allow cheating on calibration then it's trivial to achieve "quantum supremacy": simply precompute the answer and call that "calibration"!

There's no "answer" to precalculate, it's drawing from a distribution. Could you calibrate a non-quantum circuit to have a distribution passing the test? Supposedly, this is not possible even given arbitrary pre-computation time.

I can pre-calculate as many samples from the distribution as you can generate during the coherence time of the QC. Once you lose coherence, you have to re-calibrate, and during that time I can pre-compute another set of samples.

That's true, if you pre-sample you can beat the test. Is that the effect you think the calibration is having?

> Once you lose coherence, you have to re-calibrate

I didn't realize this, where did you see that?

> Is that the effect you think the calibration is having?

I have no idea. All I know (or at least I'm pretty sure of) is that you have to re-calibrate every time you lose coherence, otherwise it would not be calibration, it would be a one-time process that would be considered part of the construction of the QC.

> where did you see that?

Nowhere. I inferred it. It's possible that I'm wrong, but I'll give you long odds against. I have a pretty good understanding of how QCs work.

You can bring the complexity on a classical machine down to linear by allowing for large disk space, that is precisely the claim that IBM makes in response to the argument.

At issue, as I understand it, is whether the results from the 53-bit simulation (which would take 10,000 years, or 2.5 days, depending on who you believe) are valid or whether they are just so much junk because they haven't been calibrated and rely on calibrations that may or may not generalize. Kalai responds to Peter Shor making similar critiques in the comment threads in the blog, but it seems it comes down to clarifications on methodology, so may not be a fatal defect.

According to Scott Aaronson, Martinis himself rebutted Kalai's objections: the calibration still allows you to load any circuit later.

Reply on news.ycombinator.com