Unitary Variational Quantum-Neural Eigensolver

• The paper introduces U‑VQNHE, a hybrid variational framework that replaces non-unitary filters with a diagonal unitary phase layer to ensure normalization‐free, variationally safe energy estimation.
• The method leverages a geometric formulation on the unitary group to provide convergence guarantees and polynomial resource scaling while mitigating barren-plateau issues.
• U‑VQNHE addresses support-mismatch and statistical challenges inherent in previous DNP schemes, demonstrating reliable performance on transverse-field Ising model benchmarks.

Unitary Variational Quantum–Neural Hybrid Eigensolver (U‑VQNHE) is a hybrid variational framework for quantum ground-state estimation in which the neural component is constrained to be unitary, thereby preserving normalization by construction and maintaining the Rayleigh–Ritz variational bound. In its explicit ground-state formulation, U‑VQNHE replaces the diagonal non-unitary post-processing used in earlier variational quantum-neural schemes with a diagonal unitary phase layer acting in the measurement-record basis (Kim et al., 19 Feb 2026). In a later geometric formulation, the same framework is presented as a manifold-based design blueprint in which optimization is performed directly on the unitary group, either as a single unitary or as a product of unitary layers, with neural modules such as preconditioners, adaptive shot allocation, or learned layer selection required to respect the underlying Riemannian geometry (Qin, 27 May 2026).

1. Historical setting and motivating obstruction

The immediate predecessor of U‑VQNHE is the variational quantum-neural hybrid eigensolver (VQNHE), which augmented a shallow parameterized quantum circuit with classical neural post-processing of measurement outcomes (Zhang et al., 2021). In that line of work, the neural map was diagonal and generally non-unitary, and its purpose was to increase effective expressivity without deepening the quantum circuit. Related extensions, including VQNHE++ and other diagonal non-unitary post-processing (DNP) variants, retained the same ratio-normalized estimator structure (Zhang et al., 2021, Kim et al., 19 Feb 2026).

A later rigorous analysis identified three desiderata for such hybrids: self-contained training without prior knowledge, polynomial resource scaling, and variational consistency under finite-shot implementations (Kim et al., 19 Feb 2026). The same analysis showed that existing DNP approaches cannot satisfy these requirements simultaneously. The obstruction is statistical rather than merely algorithmic: the empirical energy is a ratio estimator whose denominator must be inferred from finite-shot samples of the bare ansatz, while the numerator is estimated from distinct measurement ensembles. If the numerator support is not contained in the sampled denominator support, the empirical objective can become ill-conditioned and even unbounded below (Kim et al., 19 Feb 2026).

A concise comparison is useful.

Scheme	Core transformation	Statistical/variational property
VQE	No neural post-processing	Standard variational upper bound
DNP / VQNHE	Diagonal non-unitary reweighting	Ratio normalization; support mismatch can make the empirical objective unbounded below
U‑VQNHE	Diagonal unitary phase layer	Normalization-free; variational safety

For DNP, the support-mismatch proposition states that if $B_M \not\subset B_a$, then there exist nonnegative weights $f$ such that the empirical objective $\widehat{E}_f$ is unbounded below (Kim et al., 19 Feb 2026). Preventing this by brute-force sampling incurs coupon-collector scaling: for i.i.d. uniform ansatz outputs on $\{0,1\}^n$, the expected number of ansatz shots needed to observe every element of a numerator support $B_M$ of size $N_M$ is

$\mathbb{E}[M_I] = 2^n H_{N_M},$

and $M_I \ge 2^n \log(N_M/\delta)$ suffices to achieve $B_M \subseteq B_a$ with probability at least $1-\delta$ (Kim et al., 19 Feb 2026). The same paper further proves that accurate ground-state reproduction by DNP generically requires an exponentially large reweighting range for constant-depth ansatzes and for unitary $f$0-design circuits, implying exponential shot complexity (Kim et al., 19 Feb 2026). U‑VQNHE was introduced precisely to remove this normalization bottleneck.

2. Formal definition of the unitary hybrid layer

For an $f$1-qubit Hamiltonian

$f$2

and a parameterized ansatz state

$f$3

U‑VQNHE replaces diagonal non-unitary filtering with a diagonal unitary

$f$4

where the neural network outputs real phases $f$5 on measurement records $f$6 (Kim et al., 19 Feb 2026). The resulting state is

$f$7

and the variational energy is

$f$8

The key theorem is variational safety: $f$9 for any normalized $\widehat{E}_f$0 and any diagonal unitary $\widehat{E}_f$1 (Kim et al., 19 Feb 2026). Because $\widehat{E}_f$2 is unitary, no normalization factor analogous to $\widehat{E}_f$3 appears.

The same construction is presented in equivalent notation as

$\widehat{E}_f$4

with

$\widehat{E}_f$5

in a scalable ground-state-estimation formulation (Kim et al., 15 Jul 2025).

The operational distinction from DNP is exact. DNP applies

$\widehat{E}_f$6

and evaluates the ratio

$\widehat{E}_f$7

whereas U‑VQNHE preserves amplitudes and modifies only phases (Kim et al., 19 Feb 2026). This phase-only restriction removes the normalization circuit, the denominator shots, and the support-mismatch failure mode.

3. Measurement transformation and training mechanics

U‑VQNHE inherits the VQNHE measurement-transformation pipeline for Pauli Hamiltonians. For each Pauli term $\widehat{E}_f$8, the circuit aggregates $\widehat{E}_f$9-parities onto a chosen star qubit $\{0,1\}^n$0 via controlled $\{0,1\}^n$1 gates from the $\{0,1\}^n$2 support of $\{0,1\}^n$3 to $\{0,1\}^n$4, then maps non-$\{0,1\}^n$5 factors to $\{0,1\}^n$6 with single-qubit rotations: $\{0,1\}^n$7 for $\{0,1\}^n$8 and $\{0,1\}^n$9 for $B_M$0 (Kim et al., 19 Feb 2026). If the final readout is $B_M$1, one defines $B_M$2, the modified string $B_M$3, and the paired string $B_M$4 obtained by flipping bits on the $B_M$5 support according to the diagonalization map (Kim et al., 19 Feb 2026).

In the unitary case, the per-term estimator becomes linear rather than ratio-normalized. Writing

$B_M$6

the estimator reported in the scalable formulation is

$B_M$7

where $B_M$8 and $B_M$9 are the output distributions of the real-part and imaginary-part measurement channels, respectively (Kim et al., 15 Jul 2025). The total energy is then

$N_M$0

An equivalent expression is given in terms of shot-normalized histograms and phase factors

$N_M$1

for the real and imaginary channels (Kim et al., 19 Feb 2026).

Several implementation consequences follow directly from this structure. First, $N_M$2 need not be physically applied on hardware; its effect can be incorporated entirely in classical post-processing of shot data via phase factors (Kim et al., 19 Feb 2026). Second, the number of measurement circuits is identical to VQE/VQNHE basis-rotation circuits up to at most a constant-factor increase; in the worst case, U‑VQNHE requires at most double circuits per Pauli term to access both real and imaginary parts (Kim et al., 19 Feb 2026, Kim et al., 15 Jul 2025). Third, because each term estimator is a bounded average of eigenvalues $N_M$3 multiplied by bounded trigonometric factors, the variance scales as $N_M$4, and Hoeffding-type concentration applies (Kim et al., 19 Feb 2026).

The optimization protocol is likewise hybrid. Quantum-ansatz parameters $N_M$5 may be updated via parameter-shift whenever generators have two distinct eigenvalues, and COBYLA is also used in reported experiments (Kim et al., 19 Feb 2026). Neural parameters $N_M$6 are updated by standard backpropagation through the classical network that outputs $N_M$7, with Adam used in the reported implementations (Kim et al., 19 Feb 2026, Kim et al., 15 Jul 2025). Both stage-wise training and alternating updates are reported as feasible; one recommended schedule is to train $N_M$8 first by VQE and then optimize $N_M$9, with $\mathbb{E}[M_I] = 2^n H_{N_M},$0 as the identity initialization (Kim et al., 15 Jul 2025).

4. Geometric formulation on the unitary group

A later geometric analysis reformulates VQE, and by explicit extension U‑VQNHE, as optimization directly over the unitary group $\mathbb{E}[M_I] = 2^n H_{N_M},$1 rather than solely over a fixed parameter vector (Qin, 27 May 2026). For a Hermitian Hamiltonian $\mathbb{E}[M_I] = 2^n H_{N_M},$2 with spectral decomposition

$\mathbb{E}[M_I] = 2^n H_{N_M},$3

and a reference state $\mathbb{E}[M_I] = 2^n H_{N_M},$4, the objective is

$\mathbb{E}[M_I] = 2^n H_{N_M},$5

The analysis also considers the ansatz-free product-unitary form $\mathbb{E}[M_I] = 2^n H_{N_M},$6 with $\mathbb{E}[M_I] = 2^n H_{N_M},$7 (Qin, 27 May 2026).

The underlying geometry is standard Lie-group geometry. The Lie algebra is

$\mathbb{E}[M_I] = 2^n H_{N_M},$8

the tangent space at $\mathbb{E}[M_I] = 2^n H_{N_M},$9 is

$M_I \ge 2^n \log(N_M/\delta)$0

and the Riemannian metric is the Frobenius inner product

$M_I \ge 2^n \log(N_M/\delta)$1

For $M_I \ge 2^n \log(N_M/\delta)$2, the Euclidean gradient is

$M_I \ge 2^n \log(N_M/\delta)$3

and the Riemannian gradient is

$M_I \ge 2^n \log(N_M/\delta)$4

Equivalently,

$M_I \ge 2^n \log(N_M/\delta)$5

and also $M_I \ge 2^n \log(N_M/\delta)$6 with $M_I \ge 2^n \log(N_M/\delta)$7 skew-Hermitian (Qin, 27 May 2026).

The Riemannian gradient descent update uses a retraction,

$M_I \ge 2^n \log(N_M/\delta)$8

with the polar retraction

$M_I \ge 2^n \log(N_M/\delta)$9

The same work notes that one can also use the exponential map

$B_M \subseteq B_a$0

or first-order Trotterizations in practice (Qin, 27 May 2026).

This formulation yields explicit convergence guarantees. In the single-unitary case, let $B_M \subseteq B_a$1 be the spectral gap to the first excited energy, where $B_M \subseteq B_a$2. If the initialization satisfies

$B_M \subseteq B_a$3

and the step size obeys

$B_M \subseteq B_a$4

then

$B_M \subseteq B_a$5

Choosing $B_M \subseteq B_a$6 gives contraction factor

$B_M \subseteq B_a$7

The same theorem uses the Lipschitz constant $B_M \subseteq B_a$8 for $B_M \subseteq B_a$9 (Qin, 27 May 2026).

The global landscape is likewise explicit. The objective has exactly $1-\delta$0 critical points $1-\delta$1 characterized by $1-\delta$2, and the Riemannian Hessian bilinear form is

$1-\delta$3

For $1-\delta$4, these are global minima, whereas every non-optimal critical point is a strict saddle; there are no spurious local minima (Qin, 27 May 2026).

The same analysis derives the gradient lower bound

$1-\delta$5

where $1-\delta$6 is the excited-space weight of $1-\delta$7. Inside the local basin $1-\delta$8, one has

$1-\delta$9

This supplies a geometric statement of when gradients vanish: namely when the state is nearly orthogonal to the ground subspace (Qin, 27 May 2026).

5. Depth, initialization, and finite-shot robustness

For product-unitary circuits,

$f$00

the Euclidean block gradient is

$f$01

and the Riemannian block gradient is

$f$02

(Qin, 27 May 2026). If the initialization satisfies

$f$03

and

$f$04

then

$f$05

so the depth-dependent contraction factor at $f$06 is

$f$07

The rate therefore deteriorates polynomially with circuit depth $f$08, rather than exponentially, and this is presented as a geometric explanation of barren-plateau behavior in deep, highly expressive unitary ansatzes (Qin, 27 May 2026).

The same work makes the depth–expressivity tradeoff explicit. If each layer lies on a $f$09-dimensional ansatz manifold $f$10, then to represent a target family $f$11 one needs

$f$12

In the universal case $f$13 for $f$14 qubits, $f$15; with hardware-efficient layers of expressivity $f$16, the depth scales as

$f$17

(Qin, 27 May 2026). The resulting flattening of the contraction factor motivates depth regularization and layerwise or blockwise training in the U‑VQNHE design.

Initialization is treated through small-angle random Pauli rotations,

$f$18

with independent $f$19 (Qin, 27 May 2026). Writing

$f$20

the theorem states that with probability at least $f$21,

$f$22

For small $f$23, $f$24 and $f$25; satisfying the convergence-basin condition requires $f$26 together with a reference state of reasonably low energy (Qin, 27 May 2026).

Finite-shot robustness is also explicit. For a Pauli decomposition

$f$27

with $f$28 shots per term and noisy estimate

$f$29

the iterates obey, with probability at least $f$30,

$f$31

Thus RGD retains the noiseless linear rate but converges only to a noise-dominated neighborhood whose radius is $f$32 (Qin, 27 May 2026).

Under non-uniform shots $f$33 with fixed total budget $f$34, the optimal allocation is

$f$35

for which

$f$36

with equality only if all $f$37 are equal (Qin, 27 May 2026). This is the basis for coefficient-adaptive shot allocation in geometry-aware U‑VQNHE.

6. Resource profile, benchmarks, and limitations

The reported resource profile is close to VQE. U‑VQNHE uses the same Pauli-term basis-rotation circuits as VQE and VQNHE, with at most double circuits per term to access real and imaginary channels, and no extra normalization circuit (Kim et al., 19 Feb 2026). Shots remain polynomial for fixed precision per term because the estimators are linear averages rather than ratio estimators (Kim et al., 19 Feb 2026). In the scalable implementation, the classical network is modest—two fully connected layers suffice in experiments—and training time remains polynomial (Kim et al., 19 Feb 2026). The geometric formulation adds explicit Riemannian-gradient and retraction computations; classically, polar retraction on $f$38 is $f$39, while on hardware the flows can be implemented via exponential maps or Trotterizations (Qin, 27 May 2026).

The empirical case for U‑VQNHE is presented primarily on transverse-field Ising models. In one reported benchmark, for a $f$40-qubit TFIM with $f$41 shots, DNP-based VQNHE rapidly drives the empirical energy below the exact ground-state energy and diverges to nonphysical magnitudes, with extreme learned weights on unmeasured strings confirming support mismatch; U‑VQNHE was introduced specifically to eliminate this pathology (Kim et al., 19 Feb 2026). In the scalable follow-up, the same instability is quantified as energies below $f$42 for $f$43-site TFIM at $f$44 shots (Kim et al., 15 Jul 2025). For $f$45-site TFIM with a two-layer hardware-efficient ansatz and $f$46 shots per circuit, U‑VQNHE consistently improves over the VQE baseline while staying within the variationally safe region; under the same conditions, DNP with bounded range $f$47 violates the variational bound and yields nonphysical energies (Kim et al., 19 Feb 2026). A separate report describes stable improvement over VQE on $f$48-site TFIM with $f$49 shots and a two-layer ansatz, and on $f$50-site TFIM with $f$51 shots, where U‑VQNHE remains within the physically valid band between the VQE energy and the exact ground-state energy (Kim et al., 15 Jul 2025).

The central limitations are equally clear. Because the learned unitary is diagonal in the computational basis, it cannot alter computational-basis amplitudes and therefore cannot emulate arbitrary amplitude filters; its effect arises through phases that interfere in rotated measurement bases for Pauli terms (Kim et al., 19 Feb 2026). U‑VQNHE depends on measurement circuits that diagonalize each Pauli term and extract both real and imaginary parts, which is standard in VQE but doubles circuits in the worst case (Kim et al., 19 Feb 2026). It also does not eliminate base-ansatz expressibility limitations: for highly nonlocal or deep ansatzes, U‑VQNHE avoids normalization issues but still relies on the underlying ansatz’s ability to reach low energies (Kim et al., 19 Feb 2026).

The open questions in the geometric program are more structural. The global landscape for product-unitary objectives can be intricate, with higher-order saddles or spurious minima; layerwise or block optimization combined with manifold-aware neural heuristics is suggested as a promising direction (Qin, 27 May 2026). Practical estimation of $f$52, $f$53, and $f$54 for tuning $f$55 and $f$56 is explicitly identified as nontrivial (Qin, 27 May 2026). Extending finite-shot robustness from energy estimation to gradient estimators, including parameter-shift settings with adaptive allocation across terms, remains open (Qin, 27 May 2026).

Taken together, these results define U‑VQNHE as a norm-preserving hybrid alternative to diagonal non-unitary post-processing, with a normalization-free estimator, exact variational safety, polynomial resource scaling at fixed precision, and a geometric theory that supplies convergence rates, initialization guarantees, finite-shot bounds, and a principled account of depth-induced trainability degradation (Kim et al., 19 Feb 2026, Kim et al., 15 Jul 2025, Qin, 27 May 2026).