acceptodds
Under review as a conference paper at ICLR 2027

Toward Fully Unitary Quantum Backpropagation

Abstract

We introduce a fully unitary quantum backpropagation framework that trains deep quantum neural networks without any mid-circuit measurement. The algorithm operates in a hybrid Schr\"odinger–Heisenberg paradigm: the forward pass alternates weight unitaries with QSVT activation layers (polynomial approximations of non-linearities such as sigmoid or tanh, realised as block-encodings), caching both pre-activation states and activation states in ancilla registers; the backward pass propagates the loss operator through both activation and weight layers in the Heisenberg picture and coherently writes gradient states into ancilla registers—entirely without measurement collapse. Three algorithmic enhancements—an operator cache, a barren-plateau detection-and-recovery mechanism, and a quantum Adam optimiser with warmup and cosine decay—are each supported by formal guarantees. Our main theoretical contributions are: (i) an per-iteration gate bound ( the gradient QSVT degree, the activation QSVT degree, the maximum layer depth, the number of layers) obtained via a quantum circuit-wrapping model that propagates the loss operator as a nested block-encoding—exact, with no Pauli-string enumeration or truncation error; (ii) an convergence rate matching classical Adam whenever the QSVT accuracy satisfies ; (iii) a probabilistically sound-and-complete barren-plateau detection guarantee for bounded gradient norm (); and (iv) a near-matching lower bound, establishing near-optimality when is polylogarithmic in . As an alternative propagation strategy we also develop, in app:pauli-mode, an approximate Pauli-propagation mode with a certified truncation-error bound . Empirically, on the MNIST digit-parity benchmark our circuit-wrapping gradient matches a state-of-the-art Mixture-of-Quantum-Experts baseline at every expert count while incurring less quantum-circuit cost per training step; and because this cost is independent of the parameter count, it affords deeper experts that lift test accuracy to —exceeding the baseline's best reported —at a per-step cost reduction the baseline cannot match.

Primary area:
learning theory
Keywords:
Quantum Machine Learning

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_zTWJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a “fully unitary” quantum backpropagation scheme for deep QNNs. The idea is a forward pass alternating weight unitaries with QSVT-implemented activations and a Heisenberg-picture propagation of the loss operator. The paper claims that a O(dL2daD+Ldan)O(dL^2d_aD + Ld_{an}) per-iteration gate bound.

Empirically, it reports MNIST-parity results against the MoQE baseline and 98.2% accuracy.

Critical strengths

The problem is very important I believe. I liked the Heisenberg-picture framing as a reasonable starting point. The paper tries to address a comprehensive study of the topic including convergence analysis, BP analysis, and complexity.

Critical weaknesses

It seems that paper is assuming the existence of a unitary nonlinear activation. QSVT applies a polynomial to the singular values of a fixed block-encoded matrix UA(l)U_A^{(l)}, which is again a linear map. Iit does not apply a sigmoid entrywise to amplitudes. It is not clear to me whether the paper considers this or not. Specially since it says in Sec. 3.2 that a circuit without inter-layer non-linearity collapses to a single unitary. So even with QSVT we still get unitary.

Each gradient component is a scalar expectation value usually. Applying a QSVT polynomial of a block-encoding of G to ∣a(l−1)⟩|a^{(l−1)}⟩ produces the vector P(G)∣a⟩P(G)|a⟩ not an expectation value ! Writing an expectation value into a register coherently is amplitude estimation, which costs O(1/ϵ)O(1/\epsilon) queries per component and must be repeated per parameter. The paper never shows how O(d)O(d) queries yield this.

The forward pass is said to cache both ∣ψ(l)⟩|\psi^{(l)}⟩ and ∣a(l)⟩=A(l)∣ψ(l)⟩|a^{(l)}⟩ = A^{(l)}|\psi^{(l)}⟩, which are two different states obtained from one copy. Holding both requires two copies. Holding all 2L cached states simultaneously requires 2L independent forward evaluations. Abbas et al. (2023) prove lower bounds on exactly this information-reuse problem; the submission must explain where its model escapes those bounds, and it does not.

Questions for authors

Q1. Is A(l)A^{(l)} a fixed unitary or state-dependent? If fixed, where is the nonlinearity? If state-dependent, how are the amplitudes of ∣ψ(l)⟩|\psi^{(l)}⟩ obtained?

Q2. Can you give the explicit circuit that maps ∣a⟩∣0⟩↦∣a⟩∣∇θ⟩|a⟩|0⟩ \mapsto |a⟩|\nabla_\theta⟩ for pl≥2p_l ≥ 2 parameters with O(d)O(d) block-encoding queries, including how the register contents are read?

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

The paper proposes a quantum algorithm for backpropagation of a quantum neural network (or parametrised quantum circuit), using QSVT to implement activation functions. The algorithm attempts to do this in a fully unitary manner without mid circuit measurements, leading to a gate complexity that is independent of the number of parameters. The algorithm uses QSVT to implement activation layers and computes the gradients by backpropagating the loss operator in the Heisenberg picture. In addition the authors propose a quantum-adam optimizer to update the circuit parameters. This is accompanied by a protocol to detect barren plateaus, a lower bound on gate complexity for the general problem and experimental validation of their results.

Critical strengths

The claimed results of this paper are certainly strong and novel. To the best of my knowledge this is the first algorithm for quantum backpropagation without needing mid-circuit-measurements, and the idea of using QSVT to implement activation functions appears novel. The construction to detect barren-plateau’s is also an interesting addition that strengthens the manuscript, as this is a common issue for variational QML approaches.

Critical weaknesses

I personally find the presentation hard to parse throughout the main text. In particular, section 4 of the manuscript would benefit greatly from additional exposition. For example, why does H^(l) only change if theta_l changes significantly? What does significantly mean in this context? In addition, much of the notation is not defined at first use. The three errors ϵapprox\epsilon_{approx}, ϵconv\epsilon_{conv} and ϵtol\epsilon_{tol} are used in the abstract without definition. LdirtyL_{dirty} is used throughout the main text and not defined until page 7. C\mathcal{C} and C(l∗)\mathcal{C}(l_*) are not defined when they first appear on line 257, and are only defined implicitly during the algorithm pseudocode.

But my biggest technical issue with the manuscript is the lack of transparency surrounding how classical data is encoded into quantum states throughout. The definitions of kets for the input data, parameter vectors and gradient vectors are not clearly defined in terms of computational basis states.

In the case of the gradient state ∣∇lθ⟩|\nabla_l^\theta \rangle this is especially confusing, since the gradient values will be expectation values of the gradient operators in equation 5. I don’t see how block encoding the gradient operator, applying QSVT and then performing a controlled application will give a state that encodes these gradients. It is also not clear what polynomial is being implemented with QSVT for the gradient computation. More detail surrounding this would help significantly.

The parameter update for quantum adam is performed using a controlled rotation, but without an explicitly stated encoding it is not clear how this would work. I believe this is possible if a digital encoding is used but does not follow trivially from a controlled rotation and I’m not sure how it is possible if an amplitude encoding is used.

Questions for authors

Can you clarify what quantum state encoding is being used throughout the manuscript, in particular please define the gradient state explicitly. If you can accompany this with a clear explanation of how applying QSVT + controlled application gives this state that would also be very helpful.

Also does the number of ancilla qubits used in this algorithm depend on the number of parameters? It seems like it does, since algorithm 1 defines W as a parameter register. If this is the case then it should be stated explicitly in the main text.

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

This paper proposes a fully unitary backpropagation framework for parameterized quantum circuits. The forward pass alternates parameterized unitaries with QSVT-based activation layers and caches intermediate states. The backward pass propagates the loss operator in the Heisenberg picture using circuit-wrapped block-encodings and attempts to prepare coherent gradient registers using QSVT, followed by coherent parameter updates. The paper presents gate-complexity, convergence, barren-plateau-detection, truncation-error, and lower-bound results. It also reports MNIST digit-parity experiments and claims parameter-independent circuit-evaluation costs and substantial improvements over parameter shift.

Critical strengths

Interesting circuit-wrapping perspective. Representing Heisenberg-propagated operators implicitly through nested circuits, rather than explicitly enumerating their Pauli expansions, is conceptually appealing and may be useful independently of the complete training framework.

The paper attempts to connect algorithm design with complexity, convergence, trainability, and resource analysis rather than presenting only a heuristic training procedure.

The authors explicitly note the absence of hardware validation, the classical state-vector nature of the experiments, and unresolved issues concerning coherent data encoding and postselection.

Critical weaknesses
  1. The central coherent gradient-preparation mechanism is not established The paper’s main algorithmic claim relies on coherently preparing a register containing the gradient components $ g_j=\langle\psi^{(l-1)}|G_j^{(l)}|\psi^{(l-1)}\rangle $. However, block-encoding Gj(l)G_j^{(l)}​ and applying QSVT does not, by itself, amplitude-encode these expectation values. Appendix T.2 essentially asserts the desired map without providing an explicit circuit or proof. The normalization, success probability, failure branches, precision, and associated resource costs are also unspecified. Consequently, the claimed parameter-independent coherent gradient computation—and the complexity results depending on it—remains unsupported.

  2. The experiments do not validate the proposed algorithm or justify the claimed cost advantage The MNIST experiments use PennyLane/Torch reverse-mode differentiation and subsequently assign analytical circuit costs to parameter shift and circuit wrapping. They do not implement the proposed QSVT activation, coherent gradient write, or coherent parameter update. Agreement between reverse-mode differentiation and parameter shift therefore does not validate the paper’s central quantum subroutines. Moreover, the reported 204.5× –408.5× improvements compare circuit-evaluation counts rather than matched gate-level resources, although a circuit-wrapped QSVT query may be substantially deeper than an ordinary PQC evaluation. The assertion that doubling expert depth is “free” also conflicts with Theorem 5.1’s dependence on circuit depth D and Appendix F’s choice d=2pmax​ .

Questions for authors

Can you provide a complete one-layer, one-parameter worked example, including an explicit circuit and input-output state calculation, demonstrating that the proposed construction prepares amplitudes proportional to ⟨ψ∣Gj∣ψ⟩\langle\psi|G_j|\psi\rangle?

What are the gradient register’s normalization, success probability, failure-branch behavior, numerical precision, and total query and gate costs? How are these costs incorporated into Theorem 5.1?

How is the input-dependent activation operator constructed, and how is this compatible with Algorithm 1 precomputing fixed activation unitaries before the forward pass?

Can you provide a small-scale executable demonstration of the actual Algorithm 1, including the QSVT activation, coherent gradient register, and coherent update, and verify its output against parameter shift?

Can you compare parameter shift and circuit wrapping using total two-qubit gates, circuit depth, ancilla requirements, precision, and success overhead rather than circuit-evaluation counts alone?

How do you reconcile the claim that deeper or more parameterized experts incur unchanged cost with the D -dependence in Theorem 5.1 and the parameter-dependent choice d=2pmax​ ?

Reviewer_f3mrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a quantum backpropagation framework combining Heisenberg-picture operator propagation, QSVT, cached intermediate states, and coherent parameter updates. It claims parameter-independent gradient computation, convergence guarantees, and substantial circuit-cost savings. The empirical study uses classical statevector training on MNIST digit parity with analytically assigned quantum costs.

Critical strengths
  1. Reducing the cost of quantum gradient computation is an important problem. Representing propagated observables through circuit composition is a useful starting point.
  2. The paper attempts to connect algorithm design, complexity analysis, and a concrete learning benchmark.
Critical weaknesses
  1. The central coherent gradient-writing step is not established. Equation (5) defines a gradient operator, but applying QSVT to that operator does not by itself produce all gradient expectation values in a separate register while leaving the input state unchanged. Sections 4.2 and T.2 need an explicit construction, input-access assumptions, and precision/success costs. Similarly, the forward pass does not explain how successive intermediate states remain cached while computation continues without copying or recomputation.
  2. The reported 204x-408x speedups are not demonstrated by an implementation of the proposed quantum training loop. Appendix A.4 uses classical reverse-mode differentiation and assigns two circuit evaluations to the proposed method. This accounting is not justified by Theorem 5.1, whose cost depends on layer count, depth, and QSVT degrees. In particular, describing doubled circuit depth as free is inconsistent with that bound.
Questions for authors
  1. Can you provide an explicit circuit for the gradient-writing and state-caching steps, including the required state-preparation access, ancillas, precision, and success probability?
  2. Can the speedup comparison be recomputed using complete gate costs for an implemented QSVT gradient routine, rather than assigning two circuit evaluations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.