---
title: 'sVQNHE: Symmetry & Sign Hybrid Eigensolvers'
url: https://www.emergentmind.com/topics/svqnhe
type: topic
---

# sVQNHE: Symmetry & Sign Hybrid Eigensolvers

Searching arXiv for sVQNHE and closely related papers.
sVQNHE is an overloaded acronym in variational quantum computing. In one usage, it denotes the **symmetry enhanced variational quantum spin eigensolver**, a symmetry-aware extension of VQE and weighted SSVQE for low-energy spectra of spin Hamiltonians. In another, it denotes the **sign-Variational Quantum-Neural Hybrid Eigensolver** or **sign-VQNHE**, a neural-guided hybrid algorithm that decouples amplitude learning from sign or phase learning. The two methods address different bottlenecks—symmetry-sector restriction in one case, and sign-structure learning plus measurement efficiency in the other—but both are formulated for NISQ-era variational workloads [2203.02444] [2507.07555].

## 1. Terminology and lineage

The acronym has been used for two distinct algorithms.

| Usage | Expansion | Core mechanism |
|---|---|---|
| 2022 usage | symmetry enhanced variational quantum spin eigensolver | Symmetry restriction through hardware-preserving circuits, cost penalties, or hybrid sector targeting |
| 2025 usage | sign-Variational Quantum-Neural Hybrid Eigensolver | Classical neural amplitude model plus shallow quantum sign or phase circuit with layerwise guidance |

The 2022 method is rooted in the VQE and SSVQE tradition. Its baseline multi-state objective is
\[
\mathrm{cost} = \sum_{i=1}^{k} w_{i} \langle \phi_{i}| U^{\dagger}(\vec{\theta}) H U(\vec{\theta}) | \phi_{i} \rangle,
\]
with \(w_1 > w_2 > \cdots > w_k > 0\), and its main intervention is to restrict the variational search to symmetry sectors relevant to the target eigenstates [2203.02444].

The 2025 method is instead a redesign of the earlier **Variational Quantum-Neural Hybrid Eigensolver** (VQNHE), where a parameterized quantum state is classically post-processed by a diagonal operator
\[
\hat f=\sum_s f(s)\,|s\rangle\langle s|.
\]
In VQNHE, the variational object is effectively
\[
|\psi_{\theta,\phi}\rangle \propto \hat f_\phi U_\theta|0^n\rangle,
\]
and the 2025 sign-VQNHE specializes this quantum-neural template to sign-structure learning by assigning amplitudes to the neural module and sign or phase information to shallow diagonal quantum layers [2112.10380].

A useful disambiguation is therefore structural rather than lexical. The 2022 sVQNHE is a **symmetry-constrained eigensolver for spin models**; the 2025 sVQNHE is a **sign-structure-focused quantum-neural hybrid eigensolver**.

## 2. Symmetry-enhanced variational quantum spin eigensolver

The 2022 sVQNHE targets low-energy excited states of spin Hamiltonians by exploiting conserved quantities to reduce the effective search space. Its primary benchmark is the antiferromagnetic Heisenberg chain
\[
H_{0} = J \sum_{i=1}^{N-1} \boldsymbol{\sigma}^{i} \cdot \boldsymbol{\sigma}^{i+1},
\]
with
\[
[H_0, S_{tot}^2]=0, \qquad [H_0, S_{\alpha}]=0 \quad (\alpha=x,y,z),
\]
where
\[
S_\alpha=\frac{1}{2}\sum_i \sigma_{\alpha}^i, \qquad S_{tot}^2 = S_x^2 + S_y^2 + S_z^2.
\]
Energy eigenstates can therefore be labeled by \((s,s_z)\), and the full Hilbert space decomposes into symmetry sectors [2203.02444].

Three symmetry-preserving strategies are introduced. In the **hardware symmetry preserving** strategy, the circuit is designed so that
\[
[U(\vec{\theta}),Q_a]=0
\]
for the relevant conserved operators \(Q_a\), and the input state is prepared in the desired sector. In the **cost-function-based symmetry preserving** strategy, symmetry is enforced by penalties added to the variational objective. In the **hybrid symmetry preserving** strategy, easy symmetries are built into the circuit and harder ones are imposed through the cost function. For Heisenberg chains, the hybrid construction preserves \(S_z\) in hardware and imposes \(S_{tot}\) through penalties.

The symmetry-preserving circuit family is built from
\[
\mathcal{N}(\theta_x,\theta_y,\theta_z) = e^{i\left(\theta_x \sigma_x^1\sigma_x^2+ \theta_y \sigma_y^1\sigma_y^2+\theta_z \sigma_z^1\sigma_z^2\right)},
\]
and phase gates
\[
P(\theta)= \begin{pmatrix} 1 & 0\\ 0 & e^{i\theta} \end{pmatrix}.
\]
For the isotropic case,
\[
\mathcal{N}_{1, 2}(\theta) = e^{i \theta (\sigma_x^1 \sigma_x^2 + \sigma_y^1 \sigma_y^2 + \sigma_z^1 \sigma_z^2)} = e^{-i\theta}(\cos{2\theta} I + i\sin{2\theta} \mathcal{P}_{1,2}),
\]
which implies
\[
[\mathcal{N}_{m,n}(\theta), S_{tot}^2]=0.
\]
Without phase gates, the ansatz preserves \(S_{tot}^2\) and \(S_z\); with phase gates, it preserves \(S_z\) but generally breaks full \(SU(2)\) spin symmetry.

For singlets, the hybrid SSVQE cost is
\[
\mathrm{cost} = \sum_{i=1,2} w_i \left( {\langle H_0 \rangle}_{i} + \beta {\langle S_{tot}^{2} \rangle}_{i}^{2} \right),
\]
with
\[
\beta=1000.
\]
For triplets, the target sector \(s=1\) implies \(S_{tot}^2=2\), and the cost becomes
\[
\mathrm{cost} = \sum_i w_i \left( {\langle H_0 \rangle}_{i} + \beta {\langle (S_{tot}^{2} - 2)^2 \rangle}_{i} \right),
\]
with numerical results reported for
\[
\beta=2.
\]
The same logic extends to other symmetries. For the transverse-field Ising chain,
\[
H_{I} = J_z \sum_{i = 1}^{N - 1} \sigma_z^i \sigma_z^{i+1} + h_x \sum_{i=1}^{N} \sigma_x^i,
\]
the relevant conserved quantity is
\[
\widetilde{S_x}=\prod_{i=1}^N \sigma_x^i, \qquad [H_I,\widetilde{S_x}]=0,
\]
and the first excited state is targeted with
\[
\mathrm{cost} = \langle H_I \rangle + \beta (\langle \widetilde{S_x} \rangle + 1)^2,
\]
using \(\beta=1\).

The conceptual content is direct: low-energy states are not generic vectors in the full Hilbert space, and excited-state preparation becomes substantially less demanding when the variational family is restricted to the correct invariant subspace. The paper’s strongest claim is that this advantage becomes more pronounced for higher excited states.

## 3. Sign-Variational Quantum-Neural Hybrid Eigensolver

The 2025 sVQNHE addresses a different bottleneck: efficient learning of nontrivial sign or phase structure in hybrid variational ansätze. Its central ansatz is
\[
\ket{\psi}=F\prod_{i=0}^{L-1}W_i G_i \ket{0},
\]
where \(F\) is a classical neural-network post-processing operator, \(W_i\) are diagonal quantum layers, and \(G_i\) are shallow non-diagonal quantum layers [2507.07555].

The defining decomposition is between amplitude and sign or phase. The neural network defines
\[
F = \sum_s f(s)\ket{s}\bra{s},
\]
with \(f(s)\ge 0\), so the classical module learns only the amplitude distribution. The quantum circuit, especially the diagonal \(W_i\), learns sign or phase structure. At layer \(l\),
\[
\ket{\psi_l} = W_l(\theta^{w_l}) G_l(\theta^{g_l}) \ket{\psi_{l-1}},
\qquad
\ket{\psi_{l,f}} = F_l \ket{\psi_l}.
\]
Energy evaluation is performed termwise for a Hamiltonian decomposition
\[
H = \sum_i c_i P_i,
\]
via
\[
E_{P_i} \gets \frac{\bra{\psi_{l,f}} P_i \ket{\psi_{l,f}}}{\braket{\psi_{l,f}}{\psi_{l,f}}},
\qquad
E_l \gets E_l + c_i \cdot E_{P_i}.
\]

The quantum sign ansatz uses shallow layers with commuting diagonal gates such as \(R_z\) and \(R_{zz}\). The paper states that diagonal gates are natural for encoding phases, that they commute with each other, and that this structure enables simultaneous measurement in one basis. The initial non-diagonal layer is often
\[
G_1 = H^{\otimes n},
\]
while later \(G_l\) may be simple \(R_y\) layers. The sign ansatz is described as employing \(R_{zz}\) gates to connect all edges of the corresponding physical model.

A second defining ingredient is **gradual transfer**. Each new \(G_l\) is initialized to approximate the previous amplitude model \(F_{l-1}\), either through the global Frobenius objective
\[
|G_l - F_{l-1}|_F^2
\]
or through KL alignment on test states,
\[
P_G^k(s) = |\langle s|G_l|\psi_{\text{test}}^k\rangle|^2, \quad
P_F^k(s) = |\langle s|F_{l-1}|\psi_{\text{test}}^k\rangle|^2,
\]
with
\[
\mathbb{E}_k\left[\mathrm{KL}(P_G^k \| P_F^k)\right]\to 0.
\]
This implements the paper’s bidirectional feedback loop: the neural network guides the next shallow quantum block, and the improved quantum phase structure feeds back into the hybrid state on which amplitude learning proceeds.

Optimization is layerwise rather than global. The parameters in \(W_l\) are updated by the parameter-shift rule,
\[
\grad_\theta f(x;\theta)=\frac{1}{2}\left[f\left(x;\theta+\frac{\pi}{2}\right)-f\left(x;\theta-\frac{\pi}{2}\right)\right],
\]
and for a parameter \(\theta_k^W\),
\[
\frac{\partial E_{P_i}(\theta^W)}{\partial \theta_k^W} =
\frac{1}{2}\left( E_{P_i}\left(\theta_k^W+\frac{\pi}{2}\right) - E_{P_i}\left(\theta_k^W-\frac{\pi}{2}\right) \right).
\]
The architecture is therefore neither a conventional fixed-depth hardware-efficient ansatz nor a purely classical neural variational state. It is a structured hybrid factorization in which amplitudes and signs are assigned to different computational modules.

## 4. Measurement models, trainability, and relation to VQNHE

The two sVQNHE variants improve different parts of the VQA stack. The 2022 method primarily reduces **search-space size**; the 2025 method primarily reduces **sign-learning burden** and **measurement cost**.

For the 2022 symmetry-enhanced method, the practical resource metric is
\[
C_R=L \times n_I,
\]
where \(L\) is the number of circuit parameters and \(n_I\) the number of optimization iterations. Hardware symmetry preserving avoids symmetry-penalty measurements but can require additional circuit synthesis and state preparation, such as the singlet product state
\[
|\psi_0\rangle = \bigotimes^{N/2} |\psi^-\rangle, \qquad
|\psi^-\rangle = \frac{|01\rangle - |10\rangle}{\sqrt{2}},
\]
which the paper states costs at least \(N/2\) extra CNOTs [2203.02444].

For the 2025 sign-VQNHE, the main resource argument is measurement compression enabled by commuting diagonal gates. Standard VQE or VQNHE training is summarized as requiring roughly \(\mathcal{O}(2d\cdot \mathrm{poly}(n))\) measurement overhead for gradients, whereas sVQNHE compresses this to \(\mathcal{O}(\mathrm{poly}(n))\). Table 1 in the paper states the measurement costs as \(\mathcal{O}(2 d n^{k'})\) for VQE, \(\mathcal{O}(2 d n^k)\) for VQNHE, and \(\mathcal{O}(n^k)\) for sVQNHE. For Hamiltonians with at most two non-\(Z\) factors per Pauli string, such as the 1D \(J_1\)-\(J_2\) model, the total number of measurements is stated as \(O(m_H)\), where \(m_H\) is the number of Pauli strings in \(H\) [2507.07555].

The 2025 paper also analyzes sampling instability through
\[
\mathrm{Var}(\hat{E}_{P_i}) =
\mathrm{Var}\left(\bra{\psi_{L,f}} P_i \ket{\psi_{L,f}}\right)\cdot \mathrm{Var}(g^{-1}(s,L)),
\]
with
\[
g(s,L) = \frac{1}{n_s}\sum_{s_i} f^2(s_i),
\qquad
\mathbb{E}[g(s,L)] = \mathbb{E}\left[f^2(s)\right],
\qquad
\mathrm{Var}(g(s,L)) = \frac{1}{n_s}\mathrm{Var}\left[f^2(s)\right].
\]
Its key asymptotic claim is
\[
\mathrm{Var}_{s \sim |\langle s | \psi_L \rangle|^2}\left(f^2(s)\right) \xrightarrow{L \uparrow} 0,
\]
so the normalization-induced variance shrinks as more amplitude structure is absorbed into the quantum part.

This line of work inherits the broader quantum-neural viewpoint of VQNHE. In VQNHE, the basic energy is
\[
\mathcal L(\theta,\phi) =
\frac{\langle 0|U_\theta^\dagger \hat f_\phi^\dagger H \hat f_\phi U_\theta|0\rangle}
{\langle 0|U_\theta^\dagger \hat f_\phi^\dagger \hat f_\phi U_\theta|0\rangle},
\]
and the paper proves that under depolarizing noise, neural-only retraining yields
\[
\delta E = O(p^2),
\]
while joint PQC and neural retraining yields
\[
\delta E = O(p).
\]
VQNHE++ extends this with a transformed Hamiltonian \(H'_\tau=W_\tau^\dagger H W_\tau\) and tri-optimization over quantum, neural, and transformation parameters [2112.10380]. The 2025 sign-VQNHE is not merely a renaming of VQNHE; it is a specialization in which the neural post-processing is constrained to nonnegative amplitudes and the quantum circuit is explicitly tasked with sign or phase modeling.

## 5. Benchmarks and reported performance

The empirical literature under the sVQNHE name is heterogeneous because the two algorithms target different tasks.

For the 2022 symmetry-enhanced method, the Heisenberg-chain ground-state benchmark at \(N=16\) compares a hardware-efficient ansatz to an \(S_z\)-conserving ansatz. With only **5 layers** of the \(S_z\)-conserving circuit, the reported fidelity is
\[
\mathcal{F}>0.95
\]
after about **300 iterations**, whereas a hardware-efficient circuit of the same depth reaches only about
\[
\mathcal{F}\simeq 0.4.
\]
The hardware-efficient circuit requires at least **15 layers** and about **2000 iterations** to exceed \(0.95\) fidelity. The reported classical resource values are
\[
C_R = 77500
\]
for the symmetry-preserving circuit at \(\mathcal{F}=0.95\), and
\[
C_R = 870400
\]
for the hardware-efficient one [2203.02444].

The same paper reports that symmetry-free SSVQE becomes increasingly difficult as the number \(k\) of targeted states grows. For the Heisenberg chain at \(N=10\), Table I gives the minimum resources required to achieve
\[
\max_{1\le i\le k} |\langle H_0\rangle_i - E_i| \le 0.5J.
\]
For \(k=1\), the values are layers \(=7\), CNOT \(=63\), \(n_I=1304\), \(C_R=208640\). For \(k=4\), they are layers \(=18\), CNOT \(=162\), \(n_I=3407\), \(C_R=1294660\). For \(k=8\), they are layers \(=24\), CNOT \(=216\), \(n_I=4205\), \(C_R=2102500\). The paper states that for \(k=8\), states \(|E_5\rangle\) to \(|E_8\rangle\) fail to converge properly even with **25 layers**. By contrast, in the hybrid symmetry-preserving setting for \(N=14\), both singlets \(|E_{S_1}\rangle\) and \(|E_{S_2}\rangle\) are reported above **0.95** fidelity after about **500 iterations** with 18 layers, and triplets \(|E_{T_1}^{(0)}\rangle\) and \(|E_{T_2}^{(0)}\rangle\) also reach
\[
\mathcal{F}>0.95
\]
with 20 layers [2203.02444].

The 2025 sign-VQNHE reports its clearest many-body result on the 6-qubit \(J_1\)-\(J_2\) model with \(J_1=1\) and \(J_2=0.6\), which the paper identifies as a regime with a nontrivial sign problem. Relative to a baseline neural network after 2000 optimization steps, sign-NN achieves
\[
R_{MAE}=-0.989,\qquad R_{Var}=-0.996,
\]
which the paper summarizes as **98.9\%** reduction in mean absolute error and **99.6\%** suppression of variance. In the same table, hea-NN gives
\[
R_{MAE}=-0.751,\qquad R_{Var}=1.955,
\]
and HEA2 gives
\[
R_{MAE}=-0.916,\qquad R_{Var}=-0.132.
\]
The average optimization steps reported in the supplementary comparison are **132.67** for sVQNHE and **1676.00** for a 2-layer HEA-VQE, with success probability **60\%** versus **5\%** [2507.07555].

On a 9-qubit 2D Heisenberg model, the same paper reports
\[
R_{MAE}=-0.872,\qquad R_{Var}=-0.832
\]
for sign-NN, versus
\[
R_{MAE}=-0.331,\qquad R_{Var}=0.554
\]
for HEA1 and
\[
R_{MAE}=-0.119,\qquad R_{Var}=33.793
\]
for hea-NN. On sign-problem-free systems, the method remains competitive. For the 9-qubit TFIM, sign-NN gives
\[
R_{MAE}=-0.640,\qquad R_{Var}=-0.728.
\]
For 9-qubit Ising,
\[
R_{MAE}=-0.416,\qquad R_{Var}=-0.455,
\]
and for 12-qubit Ising,
\[
R_{MAE}=-0.811,\qquad R_{Var}=-0.979.
\]

The MaxCut benchmarks emphasize resource efficiency under qubit-compressed encodings. On 45-vertex Erdős–Rényi graphs, the paper reports about
\[
19\% \approx \frac{1-0.84}{0.84}
\]
improvement in solution quality and
\[
85\% \approx \frac{306-46}{306}
\]
improvement in quantum resource efficiency for parameter set 2. The same section notes that sign-VQE alone shows no clear advantage over brickwork-VQE, so the observed gain is attributed to the full hybrid neural-guided strategy rather than the diagonal sign ansatz by itself [2507.07555].

## 6. Limitations, scope, and conceptual significance

The two sVQNHE variants rely on different prior structures, and their limitations follow from that reliance.

For the 2022 symmetry-enhanced method, the paper states that the approach applies only when there is prior knowledge of the relevant symmetries. Hardware symmetry preserving gives the best fidelities and the lowest optimization cost, but it can require sophisticated circuit design and nontrivial initialization. Cost-function-only symmetry preserving is more flexible, especially for nonlocal symmetries such as \(\widetilde{S_x}\), but it requires extra measurements of symmetry operators and, in some cases, higher-order expressions such as \((S_{tot}^2-2)^2\). The hybrid method is presented as the practical compromise: it captures much of the symmetry advantage while avoiding full symmetry-adapted circuit synthesis [2203.02444].

For the 2025 sign-VQNHE, the paper’s scalability claims are supported primarily by numerical experiments rather than full asymptotic proofs. Its effectiveness depends on the assumption that amplitude learning is classically manageable by the neural network, and the gradual-transfer mechanism adds classical optimization overhead. The paper states that real hardware results are not shown; the “real-world simulation” uses depolarizing gate noise and finite-shot observable estimation, with two-qubit gate error **0.005** and single-qubit gate error **0.0013**. The method is therefore positioned as scalable and robust, but not as an asymptotically proven quantum speedup [2507.07555].

Taken together, the two meanings of sVQNHE illustrate a broader trend in variational quantum algorithm design: the deliberate incorporation of **problem structure** into hybrid ansätze. In the 2022 usage, the relevant structure is **symmetry**, and the algorithm narrows the accessible Hilbert space to the physically correct sector. In the 2025 usage, the relevant structure is the **amplitude-sign decomposition** of the wavefunction, and the algorithm assigns those two components to classical and quantum modules respectively. The shared methodological thesis is that shallow NISQ circuits are more effective when they are not asked to learn an undifferentiated full-state representation.

Source: https://www.emergentmind.com/topics/svqnhe