acceptodds
Withdrawn from ICLR 2027

Identity-Paired Progressive Depth Training: When Trainability Persists Beyond Expressibility

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Variational Quantum Algorithms (VQAs) are a leading paradigm for near-term quantum computing, yet their training suffers from sensitivity to circuit depth, initialization, and landscape pathologies such as barren plateaus. We study progressive depth training (PDT)—a layerwise curriculum that trains a shallow circuit before appending new layers—and identify a fundamental obstacle: fixed entangling gates (CNOTs) in hardware-efficient ansatze cause initialization shock, an energy spike when new layers are added. We propose identity-paired progressive depth training (IP-PDT), which appends forward/inverse block pairs—each consisting of a standard rotationCNOT block followed by its reverse—that compose to the identity at initialization. Because the adjacent CNOT rings cancel, the effective circuit retains only a single entangling layer surrounded by overparameterized local rotations. We prove a simple Reachable Set Saturation Theorem: under this construction the variational manifold expands exactly once (when post-entangler rotations are first introduced) and then saturates; all subsequent depth increases provide pure overparameterization of single-qubit unitaries. Despite this saturation, progressive addition of rotation parameters can continue to improve optimization outcomes—a phenomenon we term trainability beyond expressibility. We formalize IP-PDT as a continuation method on nested manifolds, prove monotone energy guarantees under an acceptance rule, and connect energy error to ground-state fidelity through spectral-gap inequalities. A detailed resource analysis shows that IP-PDT achieves lower total gate cost than both baselines by eliminating most CNOT gates. Experiments span six-qubit Transverse-Field Ising, Tilted Ising, and Random Ising Hamiltonians, a broader nine-Hamiltonian benchmark, and system sizes up to qubits. IP-PDT matches or outperforms full-depth baselines on Hamiltonians whose ground states are well-approximated by the single-entangler reachable set, with particularly strong gains in limited-budget regimes. At the hardware-efficient baseline degrades sharply while the single-entangler methods remain trainable.

Primary area:
optimization
Keywords:
quantum, optimization, VQA, quantum circuits

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_nSmnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper focus on Progressive depth training for VQE. The authors point out that a new hardware-efficient block brings a CNOT ring that cannot be switched off, so every expansion scrambles what the optimizer just found. Their fix appends a block and its reverse, so the two rings land next to each other and cancel. That cancellation is angle-independent, so the effective circuit is one fixed CNOT ring plus a stack of local rotations. They then prove this family's reachable set stops growing at depth two, and report that training still improves afterwards.

Critical strengths
  1. Because the reachable set saturates, a depth-5 single-entangler circuit trained from scratch sees exactly the same states as IP-PDT, and comparing the two at matched step budgets separates optimization from expressibility. Most work in this area compares ansätze of different expressive power and therefore cannot attribute anything; this control appears do it well. The saturation theorem itself is also correct.
Critical weaknesses
  1. Theorem 4.7 needs φ* ≠ 0 (Remark B.2 says the homotopy is trivial at φ = 0) and Theorem 4.12 needs ‖φ*‖ = O(σ) with σ > 0. But Figure 17 shows σ = 0 matching σ = 0.05, zero initialisation matching both, and Section 5.12 concludes the jitter is not what drives anything. So the mechanism the theorems analyse is not the mechanism at work. What remains unconditional is Theorem 4.9, which is bookkeeping: accept an improvement, otherwise fall back, so the energy cannot rise. Its content is the exact fallback, and that is worth having, but it is not an explanation. The paper's central question: why does training keep helping once the manifold has stopped growing, is left unanswered by Section 4.

  2. Assumption 4.11 wants a Hessian bounded below by µ_D > 0. For this circuit that is impossible. By Lemma 4.1 a depth-D ≥ 2 state is (⊗V_q)·U·(⊗r_q)|0>. Fix any parameter point and any r'. Set V' = V·U·r·r'⁻¹·U⁻¹. Then V'·U·r'|0> = V·U·r|0>, the same state. Since the post-entangler layers reach all of SU(2)^⊗n, this is a 3n-parameter family of parameter vectors preparing one state. The energy is constant along it, so the second derivative vanishes in those directions and the Hessian has at least 3n zero eigenvalues everywhere. Hence µ_D = 0, always; at depth 1 the same argument gives n. Assumptions 4.6 and 4.11 both fail, Theorem 4.7's Jacobian is singular, and Theorem 4.12's contraction factor is undefined. What makes this awkward is that the redundancy causing it is exactly what Theorem 4.2 proves. The repair is to put the regularity on the quotient by the fibres, or to require a nondegenerate minimum on the reachable manifold itself.

Questions for authors
  1. Is it true that ∇²E_D has at least 3n zero eigenvalues everywhere for D ≥ 2? If so, how do Assumptions 4.6 and 4.11 get restated so that Theorems 4.7 and 4.12 are not vacuous?

  2. If σ = 0 performs as well as σ = 0.05, which result do you intend as the explanation of the gains?

  3. Are the Table 5 / 6 / 7 differences separate measurement streams or errors, and which initialisation produced the reported numbers?

Reviewer_2rvROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper looks at progressive depth training (PDT) for variational quantum algorithms (VQE). The authors point out that in hardware-efficient ansätze (HEA), the CNOT ring can't be "switched off”, so every time you add a layer you get an energy spike, which they call initialization shock. To fix this, they propose IP-PDT: at each expansion, append a forward block and a reverse block as a pair, so the two adjacent CNOT rings cancel and the final circuit keeps just one entangling layer plus a pile of single-qubit rotations. The authors claim IP-PDT matches or beats the baselines with 1/5 of the CNOTs, with the biggest gains at low budgets.

Critical strengths
  1. The idea is simple, and the angle is interesting.
  2. The writing is well-structured and easy to follow.
  3. The code is clean.
Critical weaknesses
  1. The novelty over prior work isn't separated out well enough. Skolik et al. (2021) and Grant et al. (2019) already tackle the problem of "new layers shouldn't disturb the current state"; they just don't handle CNOTs. The real difference in IP-PDT is pairing two blocks so the CNOTs cancel each other. That trick might well work, but the result is a final circuit with a single CNOT layer, so it reads more like "propose a cheaper ansatz, then add a curriculum of redundant rotation layers" than a genuinely new training method.
  2. I think the core theoretical results are pretty shallow, and the paper's packaging is far bigger than what the theory delivers. The headline claim is that trainability persists beyond expressibility, yet none of the theorems explain why trainability improves. All we get is an empirical observation.
  3. The experiments don't support the strong claim of "trainability beyond expressibility".
  4. Key citations are missing. For example, the key papers related to barren plateaus aren't cited.
Questions for authors
  1. Can you give any theoretical result that directly shows a trainability advantage of progressive training over training from scratch on the SE circuit? Right now Thm 3.5 only covers the case where the new layers are held fixed, which doesn't match Algorithm 1.
  2. A lot of key comparison experiments are missing. Personally, I don't think it's likely they can all be filled in during the rebuttal period.
Reviewer_PvuoOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes identity-paired progressive depth training (IP-PDT) for VQE. Adding forward/inverse blocks allows adjacent CNOT rings to cancel, preserving the previous state at the unperturbed expansion point. The resulting ansatz has one fixed entangling ring and additional local rotations. The paper proves that its reachable-state set saturates at D=2D=2, develops conditional optimization guarantees, and reports simulations on spin Hamiltonians with up to 16 qubits.

Critical strengths

1. An explicit identity-preserving growth rule. Proposition 2.1 and Algorithm 1 give a concrete construction. Before jitter is added, expansion preserves the trained state exactly. This is a useful property when growing VQE circuits.

2. Useful experimental controls and clear figures. Figure 1 illustrates the circuit construction clearly, while Figure 4 compares IP-PDT with SE circuits sharing the same reachable-state set (for D≥2D\geq2). This directly probes optimization effects without additional expressibility. The paper also reports cases where IP-PDT does not perform best.

Critical weaknesses

1. The incremental contribution and practical value of the ansatz are not convincing. Identity initialization of complete circuit blocks is already described by Grant et al. (2019, Section 3), including applications to VQE. The paper's characterization of that work as initializing only “individual gates near the identity” is inaccurate. Layerwise quantum-circuit training is also established (Skolik et al., 2021). The Introduction does not adequately distinguish IP-PDT from these prior approaches, making the claimed methodological novelty difficult to assess. IP-PDT's structural cancellation of newly added CNOT rings is a distinction, but the paper does not show why it offers a substantial benefit beyond these ideas.

Moreover, Eq. (5) and Theorem 3.2 show that every D≥2D\geq2 circuit has the form UD=(⨂qVq)Uent(⨂qrq)U_D=(\bigotimes_q V_q)U_{\mathrm{ent}}(\bigotimes_q r_q). Extra layers add local parameters but no effective entangling layers or reachable states. This limits the ground states the ansatz can represent. The fivefold CNOT reduction relative to depth-5 HEA does not establish efficiency over the equally expressive SE D=2D=2 circuit. Both have nn CNOTs, while SE D=2D=2 uses 6n6n parameters rather than 15n15n. Under the paper's unfused forward-gate-count proxy, 3T3T steps cost 21nT21nT gates for fixed SE D=2D=2, compared with 30nT30nT for IP-PDT's 1→3→51\to3\to5 schedule. The optimization benefit of the redundant parameters needs to be weighed against their cost.

2. The experimental and theoretical evidence does not yet support the claimed advantage. A matched identity-block, progressively trained baseline on the same entangling backbone is missing. Figure 3 shows copy-initialized naïve PDT tracking IP-PDT closely on some tasks. In Figure 4, SE scratch at D=2D=2 outperforms IP-PDT on TFIM at T=500T=500 (−7.425-7.425 versus −7.048-7.048). The 16-qubit results in Table 1 are mixed, with substantial variance in several cases. Section 4.5 mentions simplified ADAPT-VQE, but Figure 6 does not identify its results, and the main text does not sufficiently specify its operator pool or evaluation cost.

There are also gaps in the theory's applicability. Theorems 3.3 and 3.5 assume a positive-definite Hessian or local strong convexity in the old parameter coordinates. Yet the initial local rotations have continuous state-preparation redundancies: rqRZ(φ)∣0⟩=e−iφ/2rq∣0⟩r_qR_Z(\varphi)|0\rangle=e^{-i\varphi/2}r_q|0\rangle. At regular stationary points, these give flat Hessian directions, apparently ruling out the assumptions in the full unconstrained parameter space. In addition, Theorem 3.4's cross-stage monotonicity requires fallback to the unperturbed embedded solution, while Algorithm 1 appears to compare only against the jittered stage initialization. The analysis therefore needs a clearer connection to the implemented procedure.

Questions for authors
  1. What optimization benefit is specific to IP-PDT beyond identity-block initialization and layerwise training? Can the authors substantiate it against an equally expressive SE D=2D=2 circuit under matched evaluation and gradient costs?
  2. Can the authors clarify the simplified ADAPT-VQE implementation and provide its identifiable results and resource budget, along with a matched identity-initialized progressive control?
  3. How are the curvature assumptions in Theorems 3.3 and 3.5 compatible with continuous parameter redundancies? Does the implementation retain the unperturbed embedded solution required for Theorem 3.4?

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

Related papers

Discussion (0)

Sign in to comment.