Trainable, self-verifying variational quantum eigensolver using adiabatic warm starts
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Trainable, self-verifying variational quantum eigensolver using adiabatic warm starts".
Mira: Here's the gist of the paper. The paper introduces an adiabatic variant of the variational quantum eigensolver (AVQE) that uses iterative optimization along an adiabatic path and incorporates runtime verification to certify eigenstate accuracy and convergence to a global optimum This approach establishes a scalable, self-verifying algorithm for ground-state preparation by leveraging adiabaticity to avoid barren plateaus Pruning redundant reasoning tokens The core contribution of the work is the analysis of gradient descent tracking along an adiabatic path, showing that this method can overcome traditional VQE obstacles such as barren plateaus and local optima The authors derive conditions under which gradient-based optimization successfully prepares the adiabatic ground states They propose using energy-standard-deviation measurements at runtime to certify eigenstate accuracy and verify convergence to the global optimum This leads to a self-verifying algorithm for ground state preparation.
Kai: How the training signal is built The method involves executing VQE iteratively for a sequence of Hamiltonians along an adiabatic path Specifically, they consider a linear adiabatic schedule defined by the Hamiltonian H(λ) = (1 − λ)Hi + λHf.
Mira: The optimization is performed sequentially using warm starts, where the optimizer initializes the ansatz with parameters obtained at the previous point on the path.
Kai: For each slice t = 1 to T, they run K steps of gradient-based optimization on Eλt, starting from θt−1.
Paper summary: Kai: So, this paper, "Trainable, self-verifying variational quantum eigensolver using adiabatic warm starts," it's about taking the standard Variational Quantum Eigensolver or VQE and making it better for preparing ground states. The big idea is to use an iterative optimization process along an adiabatic path.
Mira: What that means is they’re not just running VQE once; they’re moving the Hamiltonian slowly along a path, defined by some lambda, and they optimize the parameters step-by-step, using warm starts to speed things up.
Kai: The paper claims this approach can actually avoid those tricky problems we see in standard VQE setups, specifically barren plateaus and getting stuck in local optima. They show conditions where this gradient-based optimization will successfully prepare the adiabatic ground states >
Mira: And on top of that, they propose using runtime verification by measuring energy standard deviation to certify how accurate the eigenstate is and if it’s converging to a global optimum >
Conclusion: Kai: So, looking at this paper, "Trainable, self-verifying variational quantum eigensolver using adiabatic warm starts," it seems the core contribution is building an algorithm that doesn't just find *a* good answer but verifies that the answer is actually close to the true ground state >
Mira: Exactly. The authors establish theoretical guarantees for tracking those instantaneous ground states along the adiabatic path, and they introduce a runtime verification criterion based on energy standard deviation measurements to check convergence >
Kai: What does this mean for us in practice? It means we have a scalable, self-verifying algorithm for preparing ground states that leverages adiabaticity to dodge those barren plateaus >
Mira: It suggests that if we can implement this tracking and verification method on real hardware, we get a way to generate high-fidelity ground state information more reliably >
Kai: The paper deals with the practical issues too, like how measurement noise affects the optimization dynamics and what kind of shot budget you need for those certification guarantees to hold up >
Mira: It also looks at extending this concept to when we only have an approximate representability of the quantum state, showing that verification still works even if you don't have perfect representation >
Kai: So, it’s about creating a robust framework where you not only get a result but you can prove it’s reliable using real-time measurements >
University of Ljubljana · Rudolfovo Science and Technology Centre · Quantinuum
quant-ph
Submitted: 2026-02-19
Updated: 2026-10-29
Comments: 7 pages, 2 figures + 31 pages, 5 figures
Journal ref: Phys. Rev. Research 8, L042013 (2026)
DOI: 10.1103/m538-45pk
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: The paper introduces an adiabatic variant of the variational quantum eigensolver (AVQE) that uses iterative optimization along an adiabatic path and incorporates runtime verification to certify…
Key concepts
- Adiabatic Path
- This involves slowly changing the system's Hamiltonian from an initial state H(0) to a final state H(1) over time. By following this path gradually, the system is guided toward the instantaneous ground state of each intermediate Hamiltonian, which helps avoid getting stuck in poor local minima.
- Runtime Verification
- This technique uses measurements during the optimization process to check if the variational state is actually close to a true eigenstate. Specifically, it monitors energy standard deviation; if this value falls below a certain threshold, it certifies that the state is associated with a unique eigenstate branch and meets fidelity requirements.
- Barren Plateaus Avoidance
- Barren plateaus are problems where the gradient of the cost function becomes exponentially small as the number of variational parameters grows, making optimization impossible. This method uses warm starts from previous points on an adiabatic path to ensure that optimization always starts in a region where gradients are sufficiently large and useful.
- Approximate Representability
- This considers cases where the true energy is slightly different from what the ansatz can perfectly represent. The paper shows how this error affects optimization by reducing effective constants, but crucially, it proves that the core verification guarantees remain valid even when exact representation isn't achieved.
Terminology
Summary
The paper introduces an adiabatic variant of the variational quantum eigensolver (AVQE) that uses iterative optimization along an adiabatic path and incorporates runtime verification to certify eigenstate accuracy and convergence to a global optimum This approach establishes a scalable, self-verifying algorithm for ground-state preparation by leveraging adiabaticity to avoid barren plateaus
Pruning redundant reasoning tokens
The core contribution of the work is the analysis of gradient descent tracking along an adiabatic path, showing that this method can overcome traditional VQE obstacles such as barren plateaus and local optima The authors derive conditions under which gradient-based optimization successfully prepares the adiabatic ground states They propose using energy-standard-deviation measurements at runtime to certify eigenstate accuracy and verify convergence to the global optimum This leads to a self-verifying algorithm for ground state preparation.
How the training signal is built
The method involves executing VQE iteratively for a sequence of Hamiltonians along an adiabatic path Specifically, they consider a linear adiabatic schedule defined by the Hamiltonian H(λ) = (1 − λ)Hi + λHf. The optimization is performed sequentially using warm starts, where the optimizer initializes the ansatz with parameters obtained at the previous point on the path. p. 0 For each slice t = 1 to T, they run K steps of gradient-based optimization on Eλt, starting from θt−1. p. 1 The total number of optimization updates is Nupdates = TK optimization updates. p. 2
Where the gains come from
The theoretical guarantees are established through several theorems that enable gap-conditioned, verifiable, variational ground-state tracking Theorem 1 establishes sufficient conditions for tracking the instantaneous ground state along the adiabatic path. It shows that gradient descent tracks the instantaneous ground state with Nupdates that depends polynomially on the gap ∆min and the number of variational parameters M. p. 1 Theorem 2 introduces a runtime verification criterion based on energy standard deviation, which guarantees that the variational state is associated with a unique eigenstate branch and obeys a fidelity bound of at least 8/9 given a lower bound on the spectral gap. p. 2
Measurement cost and stability under shot noise
The paper addresses the practical issue of finite measurement shots by analyzing the effect of standard shot-noise models on optimization dynamics. They show that stochastic gradients remain unbiased, and concentration bounds imply that optimization dynamics stay within the tracking region identified in Theorem 1 with high probability. p. 3 The standard-deviation-based verification remains valid in the presence of shot noise provided the measurement budget scales appropriately, specifically a number of shots per slice scaling as O(∆−4min) suffices to preserve certification guarantees with high probability. p. 3
Extension to approximate representability
The analysis is extended to include approximate representability, where Assumption 5 allows for an error term ϵ such that Eλ(¯θ(λ)) ≤ E0(λ) + ϵ. This leads to a reduced effective Polyak–Lojasiewicz constant µeff = γ∆min − ϵ and a reduced P L radius rPL = γ∆min − ϵ. p. 2 While this reduces the effective curvature, it tightens the admissible adiabatic step size through the replacement of ∆min by ∆min − ϵ. p. 3 Crucially, verification guarantees remain unchanged, as they are purely spectral and do not rely on exact representability. p. 2
Admissible range of the representability error
The representability error ϵ determines whether the variance-based certification condition can be satisfied. If the variational ansatz has a representability bias Eoptλ − E0(λ) ≤ ϵ, this bias induces a nonzero minimum achievable variance Varopt(Hλ) ≥ ∆cϵ. p. 4 Certification requires σψ(Hλ) < ∆c/2, which is satisfied whenever ϵ ≤ ∆c/4. p. 3 This establishes the upper bound on the representability error discussed in the main text. p. 6
Standard deviation propagation for a general warm start
Lemma 13 provides a covariance bound that shows how standard deviation propagates during an adiabatic step. The inequality σψ(λ + δλ) ≤ σψ(λ) + δλσψ(Hf − Hi) is derived by expanding the energy functional and applying Lemma 12 for covariance bounds. p. 3 This propagation is fundamental to proving the fidelity bounds in Corollary 2. p. 2
Total measurement cost
The total measurement cost per gradient step is Citer = 2 M K S, where S is the number of shots per Hamiltonian term. Over K gradient updates, the total measurement cost (for fixed λ) is Ctotal = 2 K M K S. p. 1 This cost scales with the shot budget S chosen according to (S7. p. 3258), leading to an approximate scaling Nupdate ≃ 128 M3∥H∥op 2∥Hf − Hi∥op 3∆min. p. 2
Comparison of P L and TDVP scalings
The three analyses presented provide complementary perspectives on the tracking dynamics. Both P L constructions yield the same overall update complexity Nupdates ∝ ∆−3min, but distribute the cost differently. p. 9 The TDVP derivation starts from a continuous-time imaginary-time flow, which exhibits the expected adiabatic scaling T ∝ ∆−2min. p. 6 For explicit discretizations, an additional factor ∆−1min appears due to the stability constraint on the update size, reflecting the stiffness of the gradient flow near small gaps. p. 2 Implicit (proximal-point) discretizations provide a complementary interpretation where each proximal update corresponds to solving a strongly convex subproblem whose contraction rate scales as µ ∼ γ∆min. p. 2 The physical behavior reflects the intrinsic adiabatic curvature of the variational manifold, requiring O(∆−1min) relaxation steps per adiabatic time scale. p. 2
Non-vanishing variance of the energy close to the minimum
Theorem 3 proves that gradients do not decrease exponentially with system size when exploring the landscape near a minimum. It shows that for perturbations δθ drawn from a normal distribution, the energy variance is polynomial in M, i. p. 8e., VarδθhE(θ∗ + δθ)i ≥ ∆6minϵ4Q124∥H∥opM7 Xk1−⟨ˆ0Pekˆ0⟩2. This result is crucial for proving the absence of barren plateaus. p. 2
Fidelity certification for the identified eigen-branch
Lemma 9 provides a bound relating standard deviation to fidelity once the relevant branch is uniquely identified. If σψ(λ) < ∆min /4, then the overlap with the actual ground state is at least 8/9. p. 3 This result allows for runtime verification to certify that the prepared state approximates the actual ground state with fidelity exceeding 8/9. p. 2
Certified evolution from P L entry and branch verification
Theorem 4 combines the verification certificate with the PL-based contraction guarantee. It states that if certain conditions are met—including overlap at least 8/9 with a unique ground-state branch, parameters inside the P L region, and step size constraints—then gradient update steps at λ+ contract toward θ⋆(λ+). p. 3
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Adaptive step size control based on runtime error: The system can dynamically adjust its adiabatic discretization step size delta lambda using a criterion like delta lambda V = Delta c / 2 - sigma psi (Hf - Hi) to ensure provable convergence to the target ground state at every stage of the adiabatic path.
-
Self-verifying eigenstate certification: The AI system can incorporate a runtime verification procedure using energy standard deviation measurements, specifically checking if sigma psi (H(λ)) is less than Delta c / 2, which certifies that the variational state is associated with a unique eigenstate branch and provides an overlap bound of at least 8/9 with the actual ground state.
-
Robustness against barren plateaus: By leveraging adiabatic tracking combined with warm starts, the system can be formulated to avoid barren plateaus because
gradients remain nonvanishing along the adiabatic path
andAvoidance of barren plateaus follows directly from adiabatic tracking and is not imposed as an independent assumption.
-
Scalable optimization complexity management: The system can utilize the derived total update scaling Nupdates <= 384 ln 2 M3 ||H||2 op ||Hf - Hi|| op Delta3 min to manage the computational cost, effectively providing a
scalable, self-verifying algorithm for ground-state computations
on NISQ hardware. -
Noise resilience in measurement: The system can maintain high probability convergence under finite shot noise by ensuring the shot budget S scales polynomially in M and inverse gap as S >= 8M sup lambda h(lambda)22 L2 H gamma4 Delta4 min 1 + q 2 log(K/delta) M2.
Sources
- Quantum Computation by Adiabatic Evolution
- A Quantum Approximate Optimization Algorithm
- Addressing hard classical problems with Adiabatically Assisted Variational Quantum Eigensolvers
- Variational-Adiabatic Quantum Solver for Systems of Linear Equations with Warm Starts
- Constant Depth Digital-Analog Counterdiabatic Quantum Computing
- A unifying account of warm start guarantees for patches of quantum landscapes
- Efficient quantum state preparation of multivariate functions using tensor networks
- Adam: A Method for Stochastic Optimization
- Warm Starts, Cold States: Exploiting Adiabaticity for Variational Ground-States
- Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak-\L{}ojasiewicz Condition
Related papers
- Fulqrum: a generalized framework for quantum subspace diagonalization
- Flux-Activated Resonant Control of a Bosonic Quantum Memory
- Machine learning the effects of many quantum measurements
- Barriers to dequantization of near-term molecular electronic structure methods
- BRIDGE: Lazy-Move Compilation for Dual-Species Neutral-Atom Quantum Computers
- Non-Hermiticity of an anomalous superradiant phase