acceptodds
Under review as a conference paper at ICLR 2027

A Trainable Closed-Form Quantum Neuron: Parameter Efficiency and Trainability at Scale

Abstract

Variational quantum neural networks (VQNNs) offer attractive theoretical properties: periodic response functions arising from quantum interference provide non-linear expressive power, and resource requirements can scale linearly with input dimension. However, training is costly and difficult to deploy, and most architectures lack an atomic learning primitive whose input–output map admits a closed form for rigorous analysis. We propose the Variational Quantum Neuron (VQN), a single-neuron-level closed-form solution that we prove is mathematically equivalent to its corresponding quantum circuit, and verify this equivalence independently via circuit simulation to within floating-point precision. This closed form enables training entirely without circuit simulation, at negligible computational cost. We show that a single VQN, using only parameters, achieves accuracy on the -bit parity task for up to , while a parameter-matched perceptron remains at chance level; matching VQN's performance with classical ReLU networks given substantially larger capacity requires over 135 times more parameters at , with this gap widening sharply as task scale increases. Our results demonstrate that a rigorously derived, quantum-circuit-equivalent closed form can serve as a parameter-efficient, readily trainable learning primitive at practical scales.

Primary area:
learning theory
Keywords:
Variational Quantum Neural Network, Quantum Neuron, Closed Form Solution

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_SwTQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper, A Trainable Closed-Form Quantum Neuron: Parameter Efficiency and Trainability at Scale, introduces a Variational Quantum Neuron (VQN) whose input–output relationship is derived analytically from a parameterized quantum circuit. Its primary contribution is an exact closed-form representation of the quantum measurement expectation, enabling classical training through automatic differentiation without repeated quantum circuit simulation. The authors further establish that the expression can be evaluated with O(m)O(m) computational complexity using only m+1m+1 trainable parameters. For binary inputs, the formulation reduces to a cosine function of a weighted linear combination, providing a compact representation of nonlinear Boolean functions such as parity. The theoretical formulation is validated by comparing the closed-form predictions with independent PennyLane simulations, yielding maximum numerical discrepancies below 7.4×10−167.4\times10^-16. Experimental evaluation focuses on mm-bit parity classification, where the VQN achieves perfect accuracy for most tested dimensions up to m=20m=20, with 97.2% accuracy at m=3m=3. Comparisons with parameter-matched perceptrons and larger ReLU networks demonstrate substantial parameter efficiency, exceeding a 135-fold parameter-count advantage at m=16m=16 relative to the tested classical configurations. However, performance deteriorates beyond m=20m=20, and the experiments primarily establish representational efficiency on synthetic parity tasks rather than generalization to practical datasets or quantum computational advantage.

Critical strengths
  1. A significant theoretical strength is the derivation of an exact closed-form input–output mapping for the proposed Variational Quantum Neuron (VQN). The authors establish its mathematical equivalence to the underlying parameterized quantum circuit and independently verify this relationship through PennyLane simulations, with numerical discrepancies below 7.4×10−167.4\times10^-16. Furthermore, the factorized formulation reduces classical evaluation complexity to O(m)O(m), enabling gradient-based optimization without repeated circuit simulation. This analytical tractability provides a mathematically well-founded connection between quantum circuit dynamics and efficiently trainable neural representations.
  2. The paper demonstrates that a single VQN with only m+1m+1 trainable parameters can represent the mm-bit parity function exactly, with an explicit parameter construction supporting this expressivity claim. The experimental results further demonstrate substantial advantages over conventional perceptrons and capacity-scaled ReLU networks. Particularly noteworthy is the m=16m=16 setting, where classical networks with more than 135 times as many parameters fail to match the VQN's accuracy under the evaluated training configurations. The combination of an explicit representational construction and systematic empirical comparisons provides meaningful evidence of parameter efficiency, making this contribution consequential to the overall assessment.
Critical weaknesses
  1. Although the closed-form derivation is mathematically well established, the representational advantage demonstrated in the paper appears to arise primarily from the periodic cosine response rather than an inherently quantum computational mechanism. For binary inputs, the proposed VQN reduces to a cosine transformation of a weighted linear combination, which can be implemented directly using a classical periodic neuron. While the authors acknowledge related work on periodic perceptrons and sinusoidal networks, the experimental comparisons are restricted primarily to conventional perceptrons and ReLU-based architectures. Without comparisons against parameter-matched sinusoidal neurons or Fourier-based models, it remains difficult to determine whether the reported parameter efficiency constitutes a substantive architectural contribution beyond existing classical periodic representations. This limitation significantly weakens the central claim of comparative expressivity.
  2. The empirical evaluation relies almost exclusively on synthetic parity classification, for which the authors explicitly construct a parameter configuration that achieves exact representation. Although this establishes theoretical expressivity, it provides limited evidence of practical trainability or general applicability. Moreover, the reported performance deteriorates substantially beyond m=20m=20, reaching 58.5% at m=24m=24 and approximately chance level at m=32m=32, despite the continued existence of exact representational solutions. The authors do not sufficiently distinguish optimization difficulties from limitations associated with sampled training data in this regime. Furthermore, the absence of experiments on continuous-valued datasets is particularly consequential because the theoretical analysis indicates potentially exponential gradient-variance decay for continuous inputs. These limitations leave the practical scalability and trainability claims insufficiently substantiated.
Questions for authors

Q1 : For binary inputs, the proposed VQN reduces to a cosine function of a weighted linear combination, suggesting a close mathematical relationship with classical sinusoidal neurons. Could the authors clarify which theoretical properties distinguish the VQN from an equivalently parameterized classical periodic neuron? In particular, would a single classical cosine neuron with the same m+1m+1 parameters and decision rule achieve identical parity classification performance? Establishing this distinction is important for assessing whether the reported parameter efficiency reflects a genuinely new learning architecture or a known property of periodic activation functions. Q2: The VQN achieves nearly perfect accuracy for m≤20m\leq20, but its performance decreases to 58.5% at m=24m=24 and approximately 50% at m=32m=32, despite the existence of analytically constructed parameters that represent parity exactly. Could the authors clarify the primary cause of this discrepancy? Specifically, have they examined whether the degradation results from optimization failure, insufficient training-set coverage, or sensitivity to parameter initialization? Distinguishing these factors would materially affect the assessment of the claimed trainability at scale. Q3: The theoretical analysis indicates that gradient variance may decay exponentially with the number of continuous-valued input features under a fixed initialization scale, whereas the experimental evaluation considers only binary inputs. Could the authors clarify whether the proposed initialization strategy or training procedure can maintain nonvanishing gradients as the number of continuous features increases? In particular, is there analytical or experimental evidence that the claimed training efficiency extends beyond the binary-input regime? This question is central to evaluating the applicability of the proposed neuron to general machine learning problems.

1 comment

We thank the reviewer for recognizing the closed form, the circuit verification, and the exact parity construction. We respond point by point below.

Q1: Would a classical cosine neuron with the same m+1 parameters achieve identical parity performance?

For binary inputs, yes: Eq. (11) and the cosine neuron are then the same function. For continuous inputs they differ: at θj=π\theta_j = \pi, VQN becomes the product unit −cos⁡(θm+1)∏jcos⁡(πxj)-\cos(\theta_{m+1})\prod_j \cos(\pi x_j), which a cosine of a weighted sum cannot represent. On soft parity (the continuous-input version of parity, with target ∏j(1−2pj)\prod_j (1 - 2p_j)) we compare against periodic and Fourier models:

mVQN (m+1 params)Cosine neuron (m+1)Sine MLP (params)Random Fourier features (features)
80.980.871.00 (321)0.87 (800)
160.960.540.89 (1,153)0.00 (≤ 1,600)
240.940.000.75 (2,497)0.00 (≤ 2,496)
320.920.000.50 (4,353)0.00 (≤ 4,352)

Median test R² over 20 seeds. For m ≥ 16, VQN outperforms all models; at m = 32, a sine MLP with 132× more parameters reaches only 0.50, and the cosine neuron and random Fourier features fail to learn.

Q2: Is the degradation beyond m = 20 due to optimization, training-set coverage, or initialization?

Mainly training-set coverage, with initialization as a second factor; the optimizer itself is not the bottleneck, since the same optimizer converges in every run when started near the exact solution. As noted in our reply to Reviewer bZ8R, m ≤ 20 and m > 20 were trained under different settings: all 2m2^m inputs for m ≤ 20, and a fixed sample of 100k–300k inputs (less than 1% of the input space) for m > 20. The revision will report the two regimes in separate tables.

  1. Coverage: switching m = 20 from all inputs to 100k samples drops accuracy from 100% to 85%; with 1M samples, accuracy at m = 24 rises from 57% to 89%.

  2. Initialization: starting near the exact solution (radius r ≤ 0.3), all 20 seeds succeed for m = 24, 28, 32; the paper’s random initialization succeeds in only 10%, 0%, 0% under the same training budget.

Q3: Can the initialization prevent vanishing gradients as the number of continuous inputs grows?

Yes. App. A.11 shows that the gradient variance at initialization factorizes over features. For binary inputs it does not decay with m; for continuous inputs at a fixed initialization scale γ\gamma it decays exponentially in the number of features. Because each feature’s contribution shrinks by O(γ2)O(\gamma^2), scaling γ2∝1/m\gamma^2 \propto 1/m keeps the variance bounded away from zero; the paper left open whether this initialization also trains well. We now confirm both points on soft-parity inputs:

mγ=1\gamma = 1γ2=1/m\gamma^2 = 1/m
40.450.35
320.410.32
2560.080.31

The measurements agree with Eq. (A.21) to within 0.5%; with γ2=1/m\gamma^2 = 1/m the variance stays at about 0.31 up to m = 256. In training, VQN initialized with γ2=1/m\gamma^2 = 1/m reaches R² 0.92 on soft parity at m = 32 (mean of 3 seeds).

Reviewer_bZ8ROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose a closed-form Variational QNN and provide a way to train them entirely on classical hardware. They showcase their approach for learning parities and compare to classical ML approaches.

Critical strengths

The authors take a specific problem that is hard to solve for their classical baselines and benchmark their approach against it. The paper is well-written and easy to follow.

Critical weaknesses

I am uncertain about the scope and impact of this work. The authors argue that the closed-form expression is the main contribution, however, it is also what makes the ML model classically tractable. It is clear that the model as it is proposed is not quantum -- as everything can and is done classically -- thus it would be a quantum-inspired classical ML model. This, however, is never acknowledged by the authors, and in the conclusion it is even stated that it is a quantum model -- without claiming advantage -- whose behaviour can be analyzed in the same form as a classical neuron can. The approach the authors propose is however to the best of my knowledge (unless I am overlooking something) inherently classical.

This by itself would not be a problem if communicated properly, however, I also think the scope of the paper is incredibly narrow. The authors outline specifically one use-case for which it outperforms the classical baselines, however, there are indeed classical models that can solve this task. The way I understand it, the authors aim to compare the performances of a classical neuron to a "quantum neuron", however, the two are fundamentally different. It is well known that QNNs rather resemble kernel methods, whereas, classical kernel methods are indeed able to solve this very task. Based on what information is given, the authors compare two fundamentally distinct quantities.

I would ask the authors for some clarification on the aspects mentioned above. In particular, (1) why is it assumed that the quantum circuit and a classical neuron are comparable quantities (in particular having in mind the connection between variational circuits and kernel methods) and (2), is your approach a quantum-inspired classical model, or what remains "quantum" about this model?

Further, the task does not really bear much practical relevance and no extensions to other types of data are acknowledged or mentioned. In fact, the only motivation that is given for the work is that training on quantum hardware or simulators is expensive, so I would be glad if the authors could provide the broader motivation or impact of this work is. In the current version of the manuscript, I am not able to pinpoint this.

Questions for authors
  • Can you give an interpretation and argue why the test accuracy of your approach degrades after 20 qubits?
1 comment

We thank the reviewer for the suggestions. We respond point by point below; all new results are on the continuous-input version of parity (soft parity).

(2) Is this a quantum-inspired classical model? What remains "quantum"?

Yes, VQN is a quantum-inspired classical model, and the revision will consistently position it as such. Its value lies in its origin: the function class is not an empirically chosen activation but is derived from a quantum circuit, and the closed form is exactly equivalent to that circuit (maximum deviation 7.4×10⁻¹⁶ from PennyLane simulation, Sec. 4.2). The circuit generalizes the parity circuit of Farhi & Neven (2018): angle encoding places the data qubits in superposition, so the readout averages the accumulated phase over all 2m2^m branches (Eq. 1). Their binary-input result is our special case Eq. 11, while for continuous inputs the circuit yields a product structure that a cosine of a weighted sum cannot represent. Such circuits are naturally suited to phase-accumulating, product-type computations such as parity, but are hard to scale, on quantum hardware or in simulation: simulating the (m+1)-qubit circuit requires 2m+12^{m+1} amplitudes (about 137 GB per input at m = 32). The closed form (Eq. 10) turns the same function class into an O(m) classical computation, so it scales directly to large m (experiments up to m = 32, gradient analysis up to m = 256), while the trained parameters can still be executed on the circuit (App. A.10).

(1) QNNs resemble kernel methods; why compare with classical neurons?

VQN is a single neuron (m+1 parameters) intended to replace a single neuron in a network, so we primarily compare it with classical neurons. We agree that the kernel view is also natural and have added this comparison. By Theorem 1, the VQN output is a linear combination of 2m2^m-dimensional features ps(x)p_s(x), with kernel k(x,x′)=∏j[(1−aj)(1−aj′)+ajaj′]k(x, x') = \prod_j [(1 - a_j)(1 - a'_j) + a_j a'_j], aj=sin⁡2(πxj/2)a_j = \sin^2(\pi x_j/2). Hence VQN is a kernel model whose coefficients are constrained to m+1 parameters. We add kernel ridge regression on soft parity (20k coefficients each):

mVQN (m+1 params)Kernel ridge, VQN feature mapKernel ridge, RBF
80.980.981.00
160.960.150.14
240.940.000.00
320.920.000.00

Median test R² over 20 seeds. For m ≥ 16, kernel ridge with unconstrained coefficients in the same feature space fails to learn, while VQN stays at 0.92 or above. Constraining the coefficients to m+1 parameters is precisely what lets VQN learn from finite data.

Practical relevance and broader motivation.

Parity and soft parity are both operations executed repeatedly in real communication systems. Soft parity is the core check-node computation in LDPC decoding (used in standards such as 5G NR, Wi-Fi, and DVB-S2): given the probabilities pjp_j that each bit is 1, it outputs ∏j(1−2pj)\prod_j (1 - 2p_j). For 0/1 inputs it is the paper's parity, i.e., the parity check used in error detection. The broader motivation is given in (2): bringing a computation structure well suited to circuits into a classical model at O(m) cost.

Q: Why does test accuracy degrade beyond m = 20?

We thank the reviewer for this question; we should have reported the two regimes separately. The results for m ≤ 20 and m > 20 were obtained under different training settings. For m ≤ 20 the model is trained on all 2m2^m inputs. For m > 20 full enumeration is infeasible (at m = 32 it would be 4.3 × 10⁹ inputs, about 1 TB in double precision), so the model is trained on a fixed random sample of 100k–300k inputs, i.e., less than 1% of the input space. The revision will report the two regimes in separate tables. The degradation reflects this change of setting rather than a limitation of the model:

  1. Training data: switching m = 20 from all inputs to 100k samples drops accuracy from 100% to 85%; increasing the data for m = 24 to 1M samples raises it from 57% to 89% (20 seeds).

  2. Initialization: starting near the exact solution (radius r ≤ 0.3), all 20 seeds succeed for m = 24, 28, 32, whereas the paper’s random initialization succeeds in only 10%, 0%, 0% under the same training budget.

This large-m learning problem is not specific to the closed form. A circuit-based model would face the same sampling constraint, and in addition its state-vector simulation is in practice limited to roughly 30–40 qubits, while quantum hardware is limited by qubit count and noise. The closed form is what makes the m > 20 regime trainable at all.

Reviewer_MuBaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a quantum "neuron," constructed as follows. For an input xx, "path probabilities" ps(x)=qs(x)2p_s(x) = q_s(x)^2 are computed for each s∈{0,1}ms \in \lbrace 0,1 \rbrace^m as products of trigonometric functions (see (4)). These path probabilities are then linearly combined into the neuron output z(x,θ)z(x,\theta), with combination coefficients parametrized by θ\theta. The paper shows that the neuron output z(x,θ)z(x,\theta) can be realized as the measurement outcome of a quantum circuit (see Figure 1), which follows from elementary quantum circuit computations. The paper then specializes to the case of binary inputs (Section 3.3) and reports experiments on the m-bit parity task, comparing against classical neural network architectures.

Critical strengths
  • The paper is well written and clearly explained.
  • Extensive experiments on the m-bit parity task are carried out; for this problem, the structure of the circuit does indeed appear favorable.
Critical weaknesses
  1. Very limited novelty relative to prior work.
  • For binary inputs (as in the m-bit parity task studied in the experiments), the quantum neuron reduces to a classical cosine neuron; see equation (11). The connection between cosine neurons and quantum circuits is well known; see equation (34) in (Farhi & Neven, 2018), which is cited in the paper. The only new element of this paper is therefore the treatment of continuous rather than binary inputs. However, all experiments are carried out with binary inputs, so the relevance of this extension is not addressed.
  • Prior work (Farhi & Neven, 2018) is also mischaracterized. It is mentioned only in the sentence On the m-bit parity task (Farhi & Neven, 2018; Cao et al., 2017) for m = 2 to 20, a single VQN with only m + 1 parameters achieves 100% accuracy, while a parameter-matched perceptron remains at chance level throughout, even though the cosine neuron already appears in that work.
  1. The experiments are purely classical.
  • The paper states: Consequently, all experiments in this paper are conducted by training the closed form of equation 9 directly in PyTorch: no quantum circuit simulation is used beyond this one-time equivalence check, so the cost of training is not tied to the cost of simulating the circuit. Hence, all experimental results concern purely classical neurons, effectively comparing a ReLU activation with a periodic cosine activation.
  • Since all experiments concern classical neurons, this further limits the novelty. As the authors themselves note, the advantage of periodic activations over other activations on the parity task has already been established: Mallick et al. (2020) prove that a three-neuron periodic perceptron can realize an arbitrary Boolean function and demonstrate this concretely on the XOR and parity tasks, outperforming standard multilayer perceptrons of matched size.
  1. Missing experimental details. The paper states that hyperparameters were selected by grid search but provides no details (e.g., the output layer and loss function for the baselines, the number of epochs, or the grid ranges). Without these details, the experimental results cannot be reproduced.
Questions for authors
  • Do the experimental findings still hold when shot noise is taken into account?
  • Could the authors provide the full experimental details (or code) required to reproduce the experiments?
  • Could the authors report baseline results with a cross-entropy loss and a longer training budget?
2 comments

We thank the reviewer for the careful reading and concrete suggestions. We respond point by point below.

W1(a): For binary inputs the VQN equals a cosine neuron; the continuous-input extension is not tested.

We agree that the two are equivalent for binary inputs. They differ for continuous inputs: the closed form is z=−A(x)cos⁡(∑jϕj(xj)+θm+1)z = -A(x)\cos\big(\sum_j \phi_j(x_j) + \theta_{m+1}\big), with nonlinear phases ϕj\phi_j and amplitude A(x)≤1A(x) \le 1; at θj=π\theta_j = \pi it becomes the product unit −cos⁡(θm+1)∏jcos⁡(πxj)-\cos(\theta_{m+1})\prod_j \cos(\pi x_j), which a cosine of a weighted sum cannot represent. We extend the paper's parity task to continuous inputs (soft parity): the input pjp_j is the probability that bit j is 1 after a noisy channel, and the target is the expected parity ∏j(1−2pj)\prod_j (1 - 2p_j); for 0/1 inputs this is m-bit parity.

mVQN (m+1 params)Cosine neuron (m+1)Best MLP (params)Kernel ridge, RBF (20k coefficients)
80.980.871.00 (321)1.00
160.960.540.89 (1,153)0.14
240.940.000.75 (2,497)0.00
320.920.000.50 (4,353)0.00

Median test R² over 20 seeds. The cosine neuron fails from m = 12 onward; for m ≥ 16, VQN with m+1 parameters outperforms all baselines.

W1(b): Farhi & Neven (2018) is mischaracterized.

We agree. Farhi & Neven (2018) already give the cosine form for binary inputs (their Eq. 34), show that θj=π/2\theta_j = \pi/2 represents parity exactly, and learn it for 6–16 bits. Our Eq. 11 coincides with their Eq. 34 up to reparameterization and a bias, so the two results confirm each other in the binary case; the revision will credit them explicitly in the Introduction, Related Work, and at Eq. 11.

Eq. 11, however, is only the binary special case of our model. Our proposal is the general closed form Eq. 1, which covers real-valued inputs and contains their function class (their Eq. 34 is Eq. 11 with the bias fixed). Our contribution goes beyond their work in three ways:

  1. Quantum-inspired and scalable. Their QNN must run on quantum hardware or in simulation, and their experiments were limited to 17-qubit simulations. Theorem 1 turns the circuit into an atomic primitive, analogous to the classical perceptron, with an exact closed-form input-output map (Eq. 1) and closed-form gradients; it is trained by O(m) classical computation (Eq. 10), so it scales directly to large m (experiments up to m = 32), while the parameters can still be executed on the circuit.

  2. Greater expressive power for real-valued inputs. Their circuit encodes inputs as computational basis states and covers only binary inputs. Our angle encoding places each data qubit in superposition, so Eq. 1 averages the accumulated phase over all 2m2^m branches; this yields a product structure (e.g., −cos⁡(θm+1)∏jcos⁡(πxj)-\cos(\theta_{m+1})\prod_j \cos(\pi x_j) at θj=π\theta_j = \pi) that their circuit and any cosine of a weighted sum cannot represent.

  3. Empirical evidence. This rebuttal adds continuous-input experiments: on soft parity, VQN with m+1 parameters outperforms all baselines for m ≥ 16 (table in W1(a)).

W2: The experiments are purely classical; the advantage of periodic activations is known.

VQN is a quantum-inspired model: its function class is derived from a quantum circuit, and the closed form is exactly equivalent to that circuit, with a maximum deviation of 7.4×10⁻¹⁶ from PennyLane state-vector simulation (double-precision round-off, Sec. 4.2). Simulating the (m+1)-qubit circuit requires 2m+12^{m+1} amplitudes (about 137 GB per input at m = 32), whereas the closed form needs O(m) operations, so it can be trained on classical hardware and scaled to large m; the trained parameters can also be executed directly on the circuit (App. A.10; shot noise in Q1). The revision will adopt this positioning explicitly.

The advantage on soft parity does not come from periodic activations: in the table above, the cosine neuron and the best MLP (a sine MLP) are both periodic models, and both fall clearly behind VQN for m ≥ 24. The difference comes from the product structure induced by the circuit.

Q1: Do the findings still hold under shot noise?

Yes. At the exact solution ∣z∣∈{0,1}|z| \in \{0, 1\}, and an input with z=0z = 0 is misclassified with N shots with probability at most 2exp⁡(−N/8)2\exp(-N/8). With the trained parameters at m = 16, accuracy is 99.7% with 30 shots and above 99.99% with 100 shots; with depolarizing noise (λ=0.5\lambda = 0.5) and a recalibrated threshold, accuracy is 99.3% with 100 shots.

W3 / Q2: Full experimental details and code.

The code for all experiments in this response, with the full training configuration of every model (output layer, loss, optimizer, hyperparameter grid, epochs, seeds), is available at https://anonymous.4open.science/r/code_7FAC/.

Q3: Baselines with cross-entropy loss and a longer training budget.

We reran the ReLU MLPs with binary cross-entropy, up to 50k steps and 10 seeds. The smallest networks reaching 100% at m = 8, 12, 16 have 481, 897, and 1,729 parameters, i.e., 53×, 69×, and 102× that of VQN (m+1 parameters).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.