acceptodds
Under review as a conference paper at ICLR 2027

Scalable Quantum Machine Learning via Multi-layer Fully-Connected Variational Quantum Circuits

Abstract

Variational quantum circuits (VQCs) face an expressivity-trainability dilemma and scalability challenges. We propose Multi-Layer Fully-Connected Variational Quantum Circuits (FC-VQC), a general-purpose quantum machine learning framework that connects local VQC blocks through measurement, deterministic parameter-free routing, and re-encoding. All trainable model parameters reside within the quantum blocks, without trainable classical neural components. We study fully connected, sliding-window, and parallel block mixing and analyze computational costs, conditional error propagation, and block information exchange. Using classical simulation, we evaluate tabular regression and classification, with our main comparison addressing spatio-temporal function approximation for the Black–Scholes, Burgers, and time-dependent oscillatory PDEs with up to spatial dimensions. Comparisons include neural-network, explicit angle-feature, tensor-network, and gradient-boosted-tree baselines. The results show that FC-VQC achieves the lowest trajectory relative MAE for all PDEs at the higher dimensions . Gradient-dynamics and depolarizing-noise experiments provide complementary empirical diagnostics of trainability and preliminary noise sensitivity.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
quantum machine learning, variational quantum circuits, scalability, expressivity, high-dimensional function approximation, PDE

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yQLZOpenReview ↗
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
  1. 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
  1. 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.

  2. 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
  1. 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.

  2. Can you add one benchmark whose target is not a function of a single linear statistic of the coordinates?

10 comments

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 R2R^2 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.

LLDNNK=3K=3K=5K=5K=7K=7K=9K=9
30.83330.88680.87730.87910.8928
50.83970.86800.85870.85910.8577
70.84860.87680.84280.86250.8483
90.82920.86240.84800.85050.8509

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 YtY_t backward in time from the known terminal payoff YT=g(XT)Y_T=g(X_T) to the initial value Y0Y_0. At each time step, the neural network takes the current stock prices XtX_t as input and estimates the hedge components Zt,iZ_{t,i} used in the BSDE. For our portfolio of individual calls, each Zt,iZ_{t,i} depends only on stock ii’s current price Xt,iX_{t,i} at that time. A sliding-window block therefore has the information it needs to learn the mapping Xt,i↦Zt,iX_{t,i}\mapsto Z_{t,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,iZ_{t,i} depends only on stock ii’s price Xt,iX_{t,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)u(t,x)=h_t(a^\top x) for a fixed vector a, its gradient would always be parallel to aa. Thus, for d>1d>1, the portfolio value cannot be expressed as a function of a single linear statistic of the stock prices.

Reviewer_yQLZOpenReview ↗

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?

Reviewer_YLCmOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper propose FC-VQC, which combine small VQC blocks into multi layer network. It is able to control local circuit width while deal with high dimensional inputs. The paper compared 3 block mixing method, including fully connected, sliding-window and parallel. Theoretically the author analyzed computational complexity, error propagation and receptive fields. In experiments, this method achieves lowest average error in tabular regression, classification and PDE benchmarks through classical simulation.

Critical strengths

None

Critical weaknesses
  1. The contribution of this paper is unclear, and the motivation is not convincing. High dimensional inputs do not necessarily require the same number of qubits, which can simply be solved by existing methods such as data re-uploading[1]. Connecting small quantum blocks is also not novel [2][3]. The paper lacks discussion and comparison with related methods.
  2. The advantage of the proposed method is unclear. Even though in some benchmark results, the proposed method has slightly lower error, the computational cost is much larger. I don't see any practical benefit of FC-VQC over classical neural networks.

[1] Pérez-Salinas et al. (2020). Data re-uploading for a universal quantum classifier. Quantum, 4, 226. DOI

[2] Alam and Ghosh (2022). QNet: A Scalable and Noise-Resilient Quantum Neural Network Architecture for Noisy Intermediate-Scale Quantum Computers. Frontiers in Physics, 9, 755139. DOI

[3] Wu et al. (2022). Scalable Quantum Neural Networks for Classification. IEEE QCE 2022, pp. 38–48. DOI

Questions for authors

None

9 comments

We thank the reviewer for their careful reading of our paper and their thoughtful comments.

Answer W1:

We agree that data re-uploading or amplitude encoding can encode high-dimensional inputs using fewer qubits. Our concern is the trade-off between expressivity and trainability. Preparing a general amplitude-encoded state can require a deep circuit, while data re-uploading increases circuit depth through successive encoding layers. Although a deeper VQC may represent more complex functions, vanishing gradients can make those functions harder to learn in practice. For high-dimensional tasks that require many re-uploading layers, this trade-off can make data re-uploading a less suitable choice.

We designed FC-VQC to address scalability and trainability together. Multiple VQC blocks encode the input features, and each block contributes its own trainable parameters. As the input dimension grows, adding blocks increases the model’s capacity without requiring deeper local circuits. We believe this additional trainable capacity helps explain why FC-VQC performed more favourably on the higher-dimensional PDE tasks. At both d=81 and d=144, an FC-VQC variant achieved lower mean trajectory error on the sampled paths than every classical baseline we tested for each of the three PDEs.

We thank the reviewer for pointing us to QNet [2] and SQNN [3]. We will cite both papers and clarify how their architectures differ from FC-VQC.

QNet [2]. The most important difference is the mapping between quantum layers. QNet uses a fixed random shuffle, so a block may receive features from only a subset of the preceding blocks. This can limit which input interactions the block can represent. Our experiments in Section 3 show that the mixing rule affects performance, and our theory in Section 4 shows how that rule determines which inputs can reach each block. Parallel blocks exchange no information, sliding-window blocks share information with their neighbours, and fully connected routing gives each block access to every preceding block in one step. Hence, the order of expressive power is FC-VQC-FC > FC-VQC-SW > FC-VQC-PL.

QNet uses a trainable classical dense layer to reduce the high-dimensional quantum features to a small number of class scores. Our concern is that, when a VQC has any trainable classical layer, that layer may account for most of the predictive performance, not VQC itself. For this reason, we designed FC-VQC to perform dimensionality reduction within the VQC blocks themselves, without any trainable classical neural network parameters.

SQNN [3]. The evaluated SQNN has one feature-extraction stage followed by a single quantum predictor. It therefore lacks multiple interacting hidden VQC layers. This limits the number of trainable parameters available to transform features in hidden layers. In FC-VQC, each additional hidden layer contributes its own trainable parameters and learns a further transformation of those shared features. SQNN also lacks structured, multistage dimensionality reduction. At a fixed circuit width, this limits the number of segments the predictor can accept and therefore the input dimension the architecture can handle. FC-VQC can reduce the feature dimension across successive output layers gradually, for example 27 -> 9 -> 3 -> 1.

Answer W2:

1 FC-VQC substantially reduces approximation error.

Tables 2–4 report relative MAE against the analytical PDE solutions on the simulated paths. For Black–Scholes at (d=144), FC-VQC-FC reduced the mean relative MAE from (0.0230) for the standard DNN to (0.0133). The absolute difference of (0.0097) represents a 42.2% reduction in the DNN’s error.

2 FC-VQC outperforms strong classical baselines on the high-dimensional PDE benchmarks.

Bowles, Ahmed and Schuld (2024) found that standard classical models generally outperformed the quantum classifiers in their benchmarking study. In our experiments, an FC-VQC variant achieved lower mean trajectory error than every tested classical model on all three PDEs at both (d=81) and (d=144). These baselines include the standard DNN, the parameter-matched DNN, and two strong gradient-boosted tree models, XGBoost and CatBoost. To the best of our knowledge, no other VQC-based model has reported such performance without support from a trainable classical neural network.

3 Computational cost on classical hardware.

We agree that FC-VQC takes longer to train on a classical computer. This is expected because the computer must simulate quantum circuits designed for direct execution on quantum hardware. We expect suitable quantum hardware to execute these circuits faster than a classical simulator. Our study focuses on the design of FC-VQC and its performance across the learning tasks we evaluated. Improving its training speed on classical computers is beyond the scope of this work.

Thank you for the detailed response. The rebuttal partially clarifies the differences between FC-VQC and existing methods such as QNet and SQNN. However, my main concerns remain unresolved. In particular, the response does not provide sufficient evidence that these architectural differences lead to meaningful improvements in trainability or practical advantages over existing methods. The computational cost also remains a concern, as the expected speedup on quantum hardware has not been demonstrated or supported by a resource analysis.

Therefore, I will maintain my original score.

We thank the reviewer for their further comments. We have now completed additional comparisons with QNet-style mixing and a SQNN-inspired baseline. FC-VQC-FC achieves lower mean trajectory error in all eight QNet-style comparisons, while our complete FC-VQC model achieves a mean test R2R^2 of 0.8928, compared with 0.7569 for the best SQNN-inspired configuration. We present these results below and clarify our computational-cost analysis.

1. Architectural differences between FC-VQC, QNet and SQNN

We acknowledge that QNet and SQNN already connect quantum circuits through measurement and re-encoding. FC-VQC combines multiple local VQC blocks, configurable hidden stages with explicit routing rules, and multi-stage dimensionality reduction using VQCs. Type 4 additionally applies deterministic input feature expansion. The feature expansion and routing introduce no trainable classical neural-network parameters.

ComponentFC-VQCQNetSQNN
Input layerMultiple local VQC blocksMultiple local VQC blocksMultiple local VQC blocks
Deterministic input feature expansionIncluded in Type 4NoneNone
Hidden layersMultiple VQC-block stagesMultiple VQC-block stagesNone
Inter-block information exchangeFully connected, sliding-window, or parallel routingFixed random shuffling, with optional classical nonlinear activationNone
Measurement and re-encodingBetween successive quantum stagesBetween successive quantum stagesBetween feature extraction and prediction
Output layerMulti-stage dimensionality reduction using VQCsClassical dense layer followed by LogSoftmaxSingle-stage dimensionality reduction using VQCs

For SQNN, the entries for hidden layers and inter-block exchange refer to additional intermediate stages: its final quantum predictor combines the features produced by the parallel extractors.

We ran additional experiments to compare FC-VQC with baselines based on QNet and SQNN, as shown below.

2. Oscillatory PDE comparison with QNet-style mixing

We implemented QNet-style nonlinear activation and random shuffling within our BSDE framework, keeping the shuffle maps fixed throughout training and evaluation. This adaptation retains our VQC circuit design and BSDE output mapping; it does not reproduce QNet's original classical classification head.

The following table compares trajectory RelMAE at the same layer counts LL and circuit depths KK. Values are mean ± standard deviation; lower is better. FC-VQC-FC results are from Tables 21 and 24.

ddLLKKQNet-style mixingFC-VQC-FCRelative error reduction
81330.0961 ± 0.00410.0816 ± 0.006615.1%
81350.1031 ± 0.00390.0818 ± 0.005920.7%
81530.0929 ± 0.00610.0830 ± 0.005910.7%
81550.0990 ± 0.00550.0826 ± 0.006316.5%
144330.1639 ± 0.00870.1004 ± 0.004938.7%
144350.1673 ± 0.03490.1148 ± 0.006231.4%
144530.1489 ± 0.00900.1059 ± 0.007828.9%
144550.1736 ± 0.04410.1163 ± 0.005333.0%

FC-VQC-FC achieves lower mean errors in all eight comparisons. Comparing the best tested configuration of each method, the reductions are 12.2% at d=81d=81 and 32.6% at d=144d=144. At each dimension, every tested FC-VQC-FC configuration also has a lower mean error than every tested QNet-style configuration. The observed advantage therefore persists across the full tested grid.

The block-mixing rule may help explain these results. FC routing gives each block access to features from every preceding-layer block, whereas SW exchanges features locally and PL keeps blocks independent. QNet's fixed random shuffling exchanges information between blocks but does not guarantee the same coverage as FC routing.

Reviewer_SmvZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FC-VQC, which composes local variational quantum circuits through measurement, fixed classical routing, and re-encoding. It studies parallel, sliding-window, and fully connected routing, alongside deterministic feature expansion for tabular tasks. The analysis covers simulation cost, conditional error propagation, receptive fields, and approximation limits arising from restricted input access. Classical simulations evaluate two tabular datasets and three PDE benchmarks with up to 144 spatial dimensions. FC-VQC variants achieve the lowest reported mean trajectory error in all six PDE settings at dimensions 81 and 144.

Critical strengths
  1. A clearly specified modular architecture with explicit resource accounting. The distinction between fixed-width SW/PL variants and the growing-width FC variant is valuable. The paper also accounts for prospective parameter-shift and measurement costs rather than equating reduced circuit width with efficient hardware training.

  2. Broad empirical comparisons and restrained theoretical interpretation. The study includes neural, explicit-feature, tensor-network, and tree baselines, with five-seed results and full configuration tables. The authors correctly distinguish structural receptive fields from actual functional dependence and acknowledge that their gradient diagnostics do not prove avoidance of barren plateaus.

Critical weaknesses
  1. The principal PDE results measure performance on training trajectories. Appendix A.4 explicitly states that the same 1,000 trajectories are used for training and evaluation, with no independent validation or test trajectories. Configurations are then selected using the reported trajectory error. These results establish fitting performance on the sampled paths, but do not establish performance on new trajectories. This substantially limits the main empirical evidence.

  2. The comparisons do not isolate the advantage of the strongest architecture. FC-DNN and FC-Angle-Linear match the three-coordinate sliding-window structure, while the strongest FC-VQC-FC models use wider local blocks and global routing. Thus, the principal gains are not compared against classical replacements with matched connectivity and local width. Moreover, the PDE solutions have separable or low-dimensional aggregate structure (Appendix A.3), so nominal input dimension alone provides limited evidence of handling general high-dimensional dependence.

Questions for authors
  1. Do the rankings in Tables 2–4 persist when configurations are selected on independent validation trajectories and evaluated on a separate, untouched test set?
  2. How does FC-VQC-FC compare with classical models using the same block widths, routing, and comparable tuning budgets?
  3. Can the authors distinguish the benefit of quantum blocks from exploiting the separable or aggregate-coordinate structure of the chosen PDEs, using suitable structure-aware classical controls?
3 comments

We thank the reviewer for their careful reading of our paper and their thoughtful comments.

Answer Q1 & W1:

We respectfully ask the reviewer to consider this design choice in light of our objective. We deliberately used the same trajectories for training and evaluation because our PDE examples are numerical function approximation tasks. Our goal is to approximate the PDE solution as accurately as possible at each time step along each simulated path. This objective differs from regression on observed data. A separate validation set is unnecessary for measuring approximation error on these fixed paths.

For example, when valuing a derivatives portfolio, a bank can price option prices by solving the Black-Scholes PDE at each time step along each Monte Carlo path of simulated market prices. In our PDE examples, we assess each model’s ability to approximate the pricing function by comparing its computed prices directly with the analytical prices on those paths without a separate validation set.

Answer Q2 & part of W2 :

We consider the standard DNN and the parameter-matched DNN the most appropriate baselines for FC-VQC-FC. The standard DNN provides a direct architectural comparison. Both models have the same multiple hidden layers with fully connected mappings between successive layers. In FC-VQC-FC, trainable VQC blocks take the place of classical neurons.

A standard DNN, however, can have a ten times as many trainable parameters as FC-VQC-FC. We therefore also include a parameter-matched DNN to compare the two architectures at the same trainable parameter budget.

We designed FC-DNN and FC-Angle-Linear specifically for comparison with FC-VQC-SW because all three use the same three-coordinate sliding-window structure. Whenever FC-VQC-SW outperforms both classical controls and FC-VQC-FC in turn outperforms FC-VQC-SW, FC-VQC-FC also outperforms both controls. A comparison with fully connected versions of FC-DNN and FC-Angle-Linear is therefore unnecessary to establish this ordering among the tested models.

Answer Q3 & part of W2 :

We designed FC-DNN to assess the contribution of the quantum blocks in FC-VQC-SW. FC-DNN replaces each VQC block with a classical neural network while preserving the same three-input, three-output block structure and sliding-window routing. For example, on the 81-dimensional Burgers PDE at (L=K=3), FC-VQC-SW achieved a mean trajectory RelMAE of 0.4136, compared with 0.5017 for FC-DNN (Table 20). Thus, with the same inputs and routing, the quantum blocks achieved lower error than their tested classical replacements.

Thank the authors for clarifying the evaluation objective and the roles of the different baselines. However, the central validation concerns remain. I therefore maintain my score at this stage.

We respectfully clarify that using the same trajectories for training and evaluation was a deliberate choice aligned with our numerical function-approximation objective. We seek to approximate the PDE solution as accurately as possible at each time step along a prescribed set of simulated paths.

Each PDE benchmark has a known analytical solution. At every simulated time–state point, we compare the ML solver’s computed value directly with the analytical value and report the trajectory RelMAE. This provides a direct, quantitative assessment of solution accuracy throughout the simulated trajectories.

For example, consider revaluing a derivatives portfolio under Black–Scholes at each time step along a fixed set of Monte Carlo market scenarios. The numerical task is to compute accurate portfolio values at those specified points. When analytical prices are available, comparing the computed prices against them directly measures the approximation error.

For this fixed-scenario objective, a separate validation set is not required to measure numerical accuracy. Evaluating another set of trajectories would assess accuracy at additional points, but it is not necessary to determine how accurately the solver approximates the known solution on the prescribed paths. We therefore ask that Tables 2–4 be interpreted as comparisons of numerical approximation accuracy on those paths.

We therefore respectfully ask the reviewer to reconsider the validation concern in light of our stated objective: numerical approximation on prescribed trajectories, with accuracy assessed directly against the analytical solutions at each simulated time–state point.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.