- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes FC-VQC to split a d-dimensional input into blocks of q coordinates, run a small variational circuit on each block, and connect the blocks across L stages through fixed, parameter-free routing of measured outputs. Trainable parameters live only inside the circuits. Three mixing rules are considered (fully connected, sliding-window, parallel), along with a cost analysis, three theoretical statements, and experiments on two small tabular tasks and three PDE/BSDE benchmarks up to d = 144, evaluated by classical state-vector simulation.
Critical strengths
- Appendix C counts things properly, and the result is more useful than the main text admits. Equation (47) shows the three mixing variants have identical rotation-parameter counts, 3dLK, so the comparison is parameter-matched by construction rather than by hand-waving, and the DNN counts in Equations (46) and (42) let a reader check the capacity claims.
Critical weaknesses
-
Appendix A.4 states it directly: "a fixed set of 1,000 Monte Carlo trajectories is used for both training and trajectory-error evaluation; no independent validation or test trajectories are used", and "the main tables report the tested configuration with the lowest mean trajectory RelMAE across these seeds". So the hyperparameters are chosen by minimising the quantity that appears in Tables 2–4. The bias is not uniform across methods, because the search grids are not: FC-VQC searches sixteen (L, K) cells at d = 9 and 36 and four at d = 81 and 144, FC-Angle-Linear searches L only, and the plain DNN searches nothing. That means the optimistic bias is largest for the model whose wins the abstract leads with. It is also large relative to the effects being claimed, in Table 11 the spread across configurations is roughly 0.82 to 0.89, while the gaps between models are 0.02 to 0.05. Please hold out a separate set of trajectories (or at minimum a disjoint random stream) for selection, report a selection-aware statistic, and show the full grid spread so a reader can see how much of each win is search.
-
Fully-connected routing uses q = ⌈√d⌉ qubits per block while sliding-window and parallel use q = 3 (Table 9), so at d = 144 the comparison is twelve-qubit local circuits against three-qubit ones. The paper concedes this ("differences involving FC cannot be attributed solely to routing") but the narrative still reads FC's wins as evidence for global mixing. The control that would fix it is cheap: sliding-window routing with q = ⌈√d⌉ and B = √d blocks, same width as FC, same parameter count, local routing. Separately, the receptive-field result cuts against the paper's own reading. Theorem 4.2 with r = 1 gives at most nine reachable blocks at d = 144 and L = 5, out of forty-eight, i.e. under a fifth of the input, and yet sliding-window lands within about a third of fully-connected (0.0175 against 0.0133 on Black–Scholes). Either the benchmarks do not need the missing coordinates, which is Weakness 2 in another form, or routing is not the operative mechanism. Theorem 4.4 makes this testable at almost no cost: estimate the support-induced floor for each variant by regressing the target on the accessible coordinates and compare against the observed errors. As written, the theory motivates the design but predicts none of the outcomes.
Questions for authors
-
Were the tabular architectures in Table 1 selected on the 15% validation split or on test R²? The reported cells are exactly the maxima of Tables 11 and 12.
-
Can you add one benchmark whose target is not a function of a single linear statistic of the coordinates?
10 comments
Discussions from Reviewer yQLZ
Thanks for your response and clarification.
The response has addressed part of my concerns. I am convinced by the explanation on Black–Scholes selection bias.
Can you please further provide a parallel (no-mixing) model with q = ⌈√d⌉, which has the same circuit width, block count, parameter count as FC, and on Burgers and the oscillatory PDE at d = 81 and 144, especially to explain whether the advantage comes from global mixing or wider circuits, so the experiment could be more clear. I would raise my rating if this experiment is provided.
Also, In the response it seems the FC–SW comparison does not isolate global routing, can you please also explain the contribution of global block exchange to FC gain?
We thank the reviewer for the careful reading and the concrete suggestions.
Answer Q1 & W1:
Concrete Strength regression and PDE experiments use different evaluation protocols, which we clarify below.
1. Regression and classification.
We split each dataset into 70% training, 15% validation and 15% test data. We trained each configuration on the training set and evaluated it on the test set. The validation split was unused. As the reviewer observed, Table 1 reports for Concrete Strength (Table 11) and Wine Quality (Table 12). The rebuttal deadline does not allow us to run the validation split.
Search grid and layer-wise comparison for Concrete Strength. Table 11 reports the full search grid for 4 layers. We compare 32t11t4t1 vs DNN. To show how the result varies across the search grid, we compare all four FC-VQC circuit depths with the DNN at each matched layer. At L={3,5,9}, FC-VQC achieved higher mean test R2 at every tested K. At L=7, it did so at K={3,7}. Overall, 14 of the 16 configurations of 32t11t4t1 exceeded the DNN. Also, whichever L was selected, FC-VQC outperformed the DNN.
2. High-Dimensional function approximation
A separate validation set is not required. We deliberately used the same 1,000 simulated trajectories for training and evaluation because it is function approximation task. The goal is to approximate the PDE solution as accurately as possible at each simulated point, which differs from regression on observed data. For example, when valuing a derivatives portfolio, a bank can price option prices by solving Black-Scholes PDE at each time step along each Monte Carlo path of simulated market prices.
Search grid. We agree that the search grids differed. For each FC-VQC variant, we tested 16 combinations of (L,K) at d={9,36} and four at d={81,144}. For Black–Scholes at d=144, mean trajectory RelMAE ranged from 0.0230 to 0.0235 for the DNN and from 0.0133 to 0.0148 for FC-VQC-FC (Table 22). Even the least accurate FC-VQC-FC configuration reduced error by 35.7% relative to the best tested DNN, and the best configuration reduced it by 42.2%. The advantage therefore held across the entire tested grid, whichever FC-VQC-FC configuration was selected.
Answer W2:
Sliding window. Under our symmetric sliding-window rule, widening the block therefore also widens the window. At d=81, the FC width gives q=9, the SW block receives inputs from itself and four neighbours on either side, covering all nine blocks. SW and FC would therefore have the same block size q=9.
Missing coordinates or global routing. We solve the Black–Scholes PDE through its BSDE formulation, in which we seek the portfolio value Yt backward in time from the known terminal payoff YT=g(XT) to the initial value Y0. At each time step, the neural network takes the current stock prices Xt as input and estimates the hedge components Zt,i used in the BSDE. For our portfolio of individual calls, each Zt,i depends only on stock i’s current price Xt,i at that time. A sliding-window block therefore has the information it needs to learn the mapping Xt,i↦Zt,i, even when the other stock prices lie outside its receptive field.
On the 144-dimensional Black–Scholes benchmark, FC achieved a mean trajectory RelMAE of 0.0133, a 24% reduction relative to SW 0.0175. FC differs from SW in two ways: it routes information globally between blocks and uses 12-qubit VQC blocks, compared with 3-qubit blocks for SW. These wider circuits allow entanglement among more qubits within each block. Both features may contribute to FC’s lower error, although this comparison does not isolate their individual effects.
Zero support-induced error. For the Black–Scholes benchmark, the support-induced error is zero for PL, SW and FC. At a fixed time t, the exact hedge component Zt,i depends only on stock i’s price Xt,i, which all three variants can access. Its conditional variance given the accessible coordinates is therefore zero, so the floor follows analytically without fitting a regression.
Answer Q2:
The Black–Scholes benchmark in Table 2 already provides such an example. At any fixed time before maturity, the portfolio value is the sum of individual call prices (Eq. 24). Changing one stock price changes only that stock’s delta, so the portfolio gradient changes direction. If u(t,x)=ht(a⊤x) for a fixed vector a, its gradient would always be parallel to a. Thus, for d>1, the portfolio value cannot be expressed as a function of a single linear statistic of the stock prices.