...(continued)Cool results! The trick in the Instantaneously-deep quantum neural networks is basically keeping the depth constant by paying a polynomial increase in the number of ancillary qubits, right?
What kind of properties does a classical distribution need to have in order to be classically hard but easy
To clarify there are two "Yifan Zhang" at Princeton ECE. One also goes by Frank. This paper is written by Yifan Zhang not Frank Zhang
...(continued)Hi KdV and Tuomas,
I am relaying this message again on behalf of Zhenghao. Thanks. Personally, I am happy that this short note is receiving some attention on scirate.
We don’t think the number of terms will go up, but we don’t have any solid evidence right now. One argument is: you can contrac
...(continued)I understand that you just add a phase-less single leg spider for post-selected measurements. However, in an end-to-end simulation, a measurement outcome can return a -1 value and would still be considered an acceptable shot if the specific detecting region/check/"closed" pauli web (defined in Boldi
...(continued)Thank you, Enrico!
We were in touch with the authors of the first paper you mention. They use quantum relative entropy as the loss function to obtain a good optimization landscape. To run their algorithm, one needs to be able to prepare Gibbs state for all the Hamiltonians traversed along the train
...(continued)Still reading the paper and it has some very interesting constructions.
I wonder if this paper https://www.nature.com/articles/s42005-024-01763-x should also be considered in the category of _generative quantum advantage_. It seems to fit "generative quantum models that are hard to simulate clas
...(continued)Post-selecting in ZX-diagrams is easy: you can add a single-leg spider (with no phase) to the qubit you are post-selecting (in the usual formalism this is like multiplying by $I \otimes \langle 0|$ or $I \otimes \langle +|$). So this doesn't add any extra terms. Sampling the measurement outcomes wou
Nice work! Could you please check if your repository has public viewing access https://github.com/franz3105/GPVQuEst. Thanks
...(continued)Hi, re: PySCF, the docstring for the selected CI module clearly states:
"This is an inefficient dialect of Selected CI using the same structure as determinant based FCI algorithm. For [an] efficient Selected CI program, [the] Dice program (https://github.com/sanshar/Dice.git) is a good candidate.
...(continued)Hi Chignons and Kristel,
I’d like to share a recent work of mine that’s closely related to yours: [Near-Optimal Parameter Tuning of Level-1 QAOA for Ising Models][1]. We show how to efficiently obtain provably optimal parameters for QAOA at $p = 1$ in linear time, without making any assumptions a
Love the title! :D
...(continued)Dear all,
Let me add a few points here.
Firstly, nice work Tim and I really appreciate the retraction of the evidence claim upon re-doing the comparison with proper HCI.Despite the upset our work (arXiv:2501.07231) of pointing out flaws in QSCI might have caused some, to date no real solu
...(continued)Dear Michal Krompiec,
Thank you for your comment.
We actually released a revision to the preprint on 10th September, in which we addressed all of the points you raise here. For example, we included HCI results generated with PyCI and found that while, yes, HCI produces the most compact wavef
...(continued)We've made a significant update for this paper. The new discovery is the transversal CCZ gates on the 5D non-Abelian self-correcting memory using a more general type of cohomology operation (classified in https://arxiv.org/abs/2411.15848) involving the higher cup product, which is able to break th
...(continued)It is surprising to see that QSCI (aka SQD) is compared (just like in the Robledo-Moreno paper) to SCI in PySCF, rather than only to the implementation of HCI in PyCI, or to HCI as implemented in the reference codes Arrow and Dice. In other words, beating PySCF's SCI does not invalidate Reinholdt's
Thanks for the comments! LCHS addresses non-unitary dynamics, and our work focuses on unitary dynamics.
...(continued)It seems that your approach relies on post-selecting the (+1) measurement outcomes. How would you simulate the post-selection or measurement process within the ZX-calculus framework? Presumably, you'd need to introduce errors into the ZX graph, then contract the graph up to the measurement nodes and
...(continued)Thank you, Joseph and Konstantinos, for sharing your really nice works! We will include them in our next paper update.
I also want to mention that our previous theory work https://arxiv.org/abs/2401.10095 has proven that all shallow quantum ML models are efficiently learnable, and are known in co
...(continued)Thanks for reading. Replying on behalf of Zhenghao.
This error model was picked over the circuit-level one, like Boldizsár Poór mentioned: "something like a ZX diagram, where there is no fixed direction of time, this noise model seems like a natural choice". We agree with this.
Operationally, it's
...(continued)Awesome work! Really nice.
I would like to bring to your attention our QNLP paper from our Oxford team [ https://arxiv.org/abs/2409.08777 ]. Of course, the learning tasks, the motivations, and the methods are different. But I claim that the Strategy for bypassing trainability issues is the same: tr
Hi guys, nice work!
In case you missed it, we also developed a method to classically train a class of quantum generative models with hardness guarantees (IQP circuits), and managed to train circuits with up to 1000 qubits.
https://arxiv.org/abs/2503.02934
...(continued)I think the $|W\rangle$ and $|CCZ\rangle$ are in different SLOCC classes, no local operations + CC can convert one to another. It would have to some non-local operations, maybe you can play with H-boxes ($|CCZ\rangle$) and the $|W\rangle$ written as a ZX-diagram, equation 7 from https://arxiv.org/ab
...(continued)Just jumping in to say that in https://arxiv.org/pdf/2506.17181 we studied this exact error model. For something like a ZX diagram, where there is no fixed direction of time, this noise model seems like a natural choice.
Regarding its relation to circuit-level noise, if you use the standard mappi
...(continued)Congrats on the results! Do you know how your noise model in Section 5 compares to standard circuit-level noise?
It seems to omit some $O(p)$ errors after two-qubit gates. On the flip side, as you explain, it has more error locations, which increases some error probabilities.
Why pick this mo
...(continued)Hi! I have a question regarding the noise model considered here. You state that "we performed numerical simulations under a circuit-level noise model", but then, when the model is described, you consider biased noise in the physical qubits and "each stabilizer measurement may independently fail with
This is great work! Looking forward to reading it carefully. Can you comment on how the algorithm complexity compares with the recent advances using Linear Combination of Hamiltonian Simulation (LCHS), such as 2508.19238?
Ah, yes, that is quite relevant! Looking back, the $RP^2$ cultivation authors cite this work, but I missed its importance. Thanks for bringing it to my attention, I'll think about how I'd want to mention it in an updated version of the paper.
...(continued)I think that is something I am looking for! Obviously to be competitive with $|T\rangle$ states we'd want a protocol that didn't use too many $|C_{XYZ}\rangle$ states, but any concrete protocol would be a great place to start. Optimizations can come later.
Although after twenty minutes of fiddlin
...(continued)Congrats on the nice cultivation results, Jahan! I also wanna bring your attention to the work that first realized logical Hadamard on $RP^2$ surface code: https://journals.aps.org/prxquantum/abstract/10.1103/PRXQuantum.5.020360, which is the type of gate that can be measured for your cultivatio
...(continued)Hi Jahan, using three pi/3 phase states we can produce a |W> state. So it looks like with the pi/6 states that protocol produces you could construct a |W> state using 6 of them (though there might be a smarter way than just combining two pi/6 together to get a pi/3).
Though I don't have the details
I, the author, have found out some bugs regarding figures that should be fixed in the next version soon.
Good to know 😌. Thanks!
...(continued)I think it's wrong.
Shor's algorithm produces states that have exponentially huge numbers of likely outputs. Small subsections of the output will be exponentially close to uniformly random. So sampling small parts of the output is equivalent to sampling a uniformly randomly generated number. The
...(continued)I agree! Being able to (even probabilistically) produce a $|CCZ\rangle$ state or similar with $|C_{XYZ}\rangle$ states would basically solve the problem.
I started out restricting to $CX$ gates and $e^{i\pi Z/12}$ rotations (done probabilistically with $|C_{XYZ}\rangle$ states as specified in Br
...(continued)The main reason to expect the X+Y+Z magic state to be good is that it reaches further outside of the stabilizer octohedron compared to the X+Y magic state. It is "more magical". But conversions between different forms of magic are often lossy... so it could still be that it ends up more expensive to
...(continued)Using this space to advertise a question I have: are $|C_{XYZ}\rangle$ magic states (i.e., eigenstates of $(X+Y+Z)$) useful? Because while $|T\rangle$ states are kind of hard to cultivate on the surface code, $|C_{XYZ}\rangle$ states are quite easy, see [my thread here][1].
But just because they
@craig-gidney Thoughts?
We have updated the paper
with two new cultivation types.
Hope you'll enjoy.
More interestingly, the first author "Scited" **every single paper** listed here on Scirate but did not "cite" in their paper : P
...(continued)Hi, Alex
We thank you for pointing out the references. Indeed, we find Theorem 1.2 of https://arxiv.org/pdf/2505.16715 to be identical to our result; however, the proof differs somewhat, particularly in its dependence on the operator norm and accuracy.
Regarding the second paper, the BQP-hard
...(continued)To the authors,
Since last year, there has been substantial progress on trace estimation of quantum state powers, and I have been following this ongoing line of research with interest.
I have a few concerns regarding your manuscript. First, your main result, Theorem 1, is **completely identica
Congrats on your paper! Is your data available? Would you consider making your source code public?
...(continued)I just wanted to point out a possible typo in the no-distillation proof that would slightly change the exponential scaling required in the number of copies (but of course not its consequence).
If I'm not mistaken, the definition of $m_j $ (D11) implies $\sum_j m_j = d$, leading to $\sum_i \delta_i^
Nice use (and mention) of ChatGPT in research work :)
...(continued)Hi Stergios, Mark and Joschka,
Thanks for the response! Empirically Tesseract does have polynomial scaling - here is a log-log plot showing that Tesseract has comparable runtime to belief propagation (without LSD) for SI1000 superdense color code circuits:
Braneshop, and
the US National Science Foundation.
SciRate