Hamiltonian Variational Ansatz (HVA)

• Hamiltonian Variational Ansatz (HVA) is a structured variational circuit design that builds layered unitaries from decomposed Hamiltonian terms, ensuring physical relevance.
• It preserves key symmetries with a reduced parameter count, enabling tailored simulation of ground states, critical phenomena, and excitation spectra.
• Empirical studies show HVA mitigates barren plateaus and balances controlled expressivity with optimization, improving state fidelity across various quantum models.

Searching arXiv for recent and foundational papers on Hamiltonian Variational Ansatz to ground the article in current literature. Searching arXiv for recent and foundational papers on Hamiltonian Variational Ansatz to ground the article in current literature. The Hamiltonian Variational Ansatz (HVA) is a problem-inspired variational circuit family in which the generators are taken directly from a decomposition of the target Hamiltonian. In its standard form, one writes a Hamiltonian as $H=\sum_s H_s$ and constructs a layered unitary $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$, with $\ket{\psi_0}$ chosen as a simple reference state, often the ground state of one component of the Hamiltonian. This places HVA between adiabatic state preparation and QAOA: it is a variationally reparameterized Trotterization of physically meaningful time evolution rather than a generic hardware-efficient circuit. Across recent literature, HVA has been used for ground-state preparation, critical-state simulation, excitation spectroscopy, lattice gauge theory, and measurement-based implementations (Wiersema et al., 2020, Peng et al., 18 Jul 2025).

1. Formal construction and canonical interpretation

In the broadest formulation, HVA starts from a decomposition of the target Hamiltonian into non-commuting pieces,

$H=\sum_s H_s,$

and defines a depth-$p$ variational state by alternating exponentials of those pieces,

$$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$

This is the form analyzed in detail by Wiersema et al., who emphasize that HVA is a Hamiltonian-structured generalization of QAOA: instead of a single cost term and a single mixer, one may use several Hamiltonian components $H_s$, each with its own parameter in each layer (Wiersema et al., 2020).

The canonical intuition is Trotterized adiabatic or real-time evolution. For the 1D transverse-field Ising model (TFIM), recent work uses

$$H=-J\sum_{j=1}^N \sigma_j^z\sigma_{j+1}^z-h\sum_{j=1}^N \sigma_j^x,$$

with $J=h=1$ at criticality, and a standard HVA layer

$$U_{\rm HVA}(\beta_l,\gamma_l) = \prod_{j=1}^N \exp\!\Bigl(-i\frac{\beta_l}{2}Z_jZ_{j+1}\Bigr) \prod_{j=1}^N \exp\!\Bigl(-i\frac{\gamma_l}{2}X_j\Bigr),$$

acting on the initial state $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$0. The corresponding VQE objective is the energy expectation $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$1. In this construction, each layer is explicitly interpreted as a short-time evolution under the interaction term followed by the transverse-field term (Du et al., 26 Apr 2026).

The same logic extends beyond spin chains. For the quantum Rabi model, HVA is built from

$$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$2

with blocks

$$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$3

so that the full ansatz is a product of $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$4 such blocks acting on $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$5. In the Hubbard-chain setting, the same adiabatic-Trotter logic leads to a variational Hamiltonian ansatz built from exponentials of the on-site interaction and the even- and odd-bond hopping operators. These constructions differ in microscopic content, but they share the same principle: variational parameters play the role of optimized evolution times under physically meaningful Hamiltonian terms (Peng et al., 18 Jul 2025, Martin et al., 2021).

2. Symmetry, parameterization, and expressivity

A defining structural property of HVA is that it inherits the symmetries of the chosen Hamiltonian decomposition. If each generator $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$6 commutes with a symmetry operator and the reference state lies in a fixed symmetry sector, the full ansatz remains confined to that sector. For the TFIM, the global $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$7 parity

$$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$8

commutes with both the $$\ket{\psi_p(\boldsymbol{\theta})}=\prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}$$9 and $\ket{\psi_0}$0 terms, so a standard HVA initialized in a definite-parity state preserves parity exactly. In finite systems this makes HVA naturally target the symmetric ground state, including cat-like states in low-field regimes. The same theme appears in the transverse-field cluster model, where symmetry-preserving HVA remains confined to a fixed $\ket{\psi_0}$1 sector (Tripathi et al., 19 Feb 2026, Park, 2021).

This symmetry preservation is closely tied to HVA’s low parameter count. In translationally invariant TFIM realizations, a depth-$\ket{\psi_0}$2 HVA uses only $\ket{\psi_0}$3 parameters, or $\ket{\psi_0}$4 in the symmetry-breaking extension HVA-SB. By contrast, hardware-efficient circuits such as EfficientSU2 use a number of parameters proportional to system size times depth. The literature repeatedly stresses that this reduced parameterization is not merely economical: it encodes an inductive bias toward Hamiltonian-relevant submanifolds of Hilbert space (Tripathi et al., 19 Feb 2026, Tripathi et al., 22 Apr 2026).

The corresponding trade-off is reduced global expressivity. Using the $\ket{\psi_0}$5 frame potential, TFIM studies find that EfficientSU2 is the most expressive, HVA the least expressive, and HVA-SB intermediate. The restricted manifold can be beneficial for ground-state problems, but it can also prevent access to symmetry-broken states or other sectors unless extra layers are introduced. In the symmetry-breaking construction proposed by Park, additional $\ket{\psi_0}$6 layers explicitly violate the relevant $\ket{\psi_0}$7 symmetry, thereby allowing access to parity-violating states and substantially shortening the depth needed for symmetry-broken phases (Tripathi et al., 19 Feb 2026, Park, 2021).

A related design choice appears in momentum-space excited-state VQE for the XXZ chain. There, the implemented HVA preserves translation symmetry and fermion parity exactly, so momentum $\ket{\psi_0}$8 is conserved throughout optimization, but it does not preserve $\ket{\psi_0}$9 particle number because the ansatz alternates separate $H=\sum_s H_s,$0, $H=\sum_s H_s,$1, and $H=\sum_s H_s,$2 bond rotations. The paper argues that for quasiparticle spectroscopy this symmetry allocation is advantageous: preserving momentum is more important than preserving particle number exactly, because energetic optimization suppresses leakage into undesired sectors (Velury et al., 20 Nov 2025).

3. Trainability, barren plateaus, and overparameterization

A major line of HVA research concerns trainability. Wiersema et al. report that HVA exhibits favorable structural properties relative to hardware-efficient ansätze, including mild or entirely absent barren plateaus for TFIM and substantially milder gradient decay than random circuits for XXZ. They also show that initialization matters: for XXZ, random initialization can lead into highly entangled, random-like regions of the ansatz manifold, whereas identity initialization keeps optimization in a structured subspace and yields fidelities above $H=\sum_s H_s,$3 (Wiersema et al., 2020).

A more formal resolution is developed by Park and Killoran. They show that when HVA parameters are constrained so that the circuit is well approximated by time evolution under a local Hamiltonian, gradients do not become exponentially small with system size. Their theorem yields a short-time regime in which gradients remain $H=\sum_s H_s,$4, and they propose a constrained initialization with per-block total angle $H=\sum_s H_s,$5. They further introduce a repeated constrained ansatz that preserves the non-barren-plateau behavior while improving expressivity. This result is specific to local Hamiltonian structure and does not generalize to arbitrary parameterized circuits (Park et al., 2023).

Optimization also displays an overparameterization phenomenon. In TFIM and XXZ, Wiersema et al. observe a size-dependent computational phase transition: below a threshold depth $H=\sum_s H_s,$6, convergence is sensitive to initialization and local minima are common; above it, optimization becomes almost trap free, and the fraction of successful runs rapidly approaches 1. The threshold appears to scale at most polynomially in system size for the models studied, in contrast to the exponential thresholds associated with learning generic Haar-random unitaries (Wiersema et al., 2020).

Noise modifies this picture. A benchmark of eight optimizers for truncated HVA in quantum chemistry finds that gradient-based methods perform best in noiseless statevector simulations, while population-based methods, especially CMA-ES, are more resilient under sampling noise. Hartree–Fock initialization reduces the number of function evaluations by $H=\sum_s H_s,$7–$H=\sum_s H_s,$8 and consistently improves final accuracy; beyond approximately $H=\sum_s H_s,$9 shots, the study reports diminishing returns because a sampling-noise precision floor dominates further improvement (Illésová et al., 28 May 2025). This suggests that HVA trainability is a joint property of ansatz geometry, initialization, optimizer class, and noise model rather than of circuit structure alone.

4. Entanglement, geometry, and dynamical interpretation

Recent work has shifted attention from static entanglement magnitude to the dynamics of entanglement within HVA. Using the Fubini–Study geometry of projective Hilbert space, one study defines the geodesic distance

$p$0

and the geometric-phase fraction

$p$1

In their TFIM comparison, hardware-efficient ansätze exhibit $p$2 almost everywhere, indicating evolution dominated by geometric phase. HVA behaves differently: early layers and optimized intermediate layers show reduced $p$3, which the authors interpret as enhanced dynamical-phase contribution aligned with the problem Hamiltonian (Du et al., 26 Apr 2026).

Within the same framework, entanglement is quantified by the bipartite von Neumann entropy of a half-system partition,

$p$4

The key empirical finding is differential rather than absolute: in HVA, larger layer-to-layer changes in entanglement entropy correlate positively with larger reductions in geodesic distance toward the target ground-space manifold, whereas no persistent correlation is found in hardware-efficient circuits. This is the sense in which the paper states that “more entanglement consumption correlates directly with faster quantum state evolution” for HVA (Du et al., 26 Apr 2026).

A complementary TFIM study reaches a different but compatible conclusion about entanglement fidelity. In 1D, pure HVA reproduces the finite-size symmetric entanglement structure, including large single-site entropy in the ordered phase for a 10-site chain, while EfficientSU2 and HVA-SB tend to find symmetry-broken lower-entropy states. In 2D, HVA can perform poorly in low-entanglement regimes at large transverse field despite the simplicity of the true ground state, which the authors attribute to limited flexibility and optimization difficulty. This suggests that HVA’s entanglement behavior is faithful to the target physics when its symmetry sector is appropriate, but fidelity and trainability need not coincide across regimes (Tripathi et al., 19 Feb 2026).

Entanglement also appears as a constructive mechanism in the quantum Rabi model. There, HVA progressively transforms an initial bosonic vacuum into a critical squeezed state: Wigner functions become increasingly elliptical with depth, only even Fock states are populated, and quadrature variances $p$5 and $p$6 saturate in a manner consistent with minimum-uncertainty squeezing. The authors interpret each HVA block as contributing a roughly fixed squeezing factor, which is why the required depth scales linearly with the effective system size $p$7 (Peng et al., 18 Jul 2025).

5. Major realizations and extensions

HVA has been instantiated across a wide range of models. In 1D spin systems, the TFIM and XXZ chains remain canonical benchmarks. Wiersema et al. report that for TFIM and XXZ ground states, depths $p$8 are sufficient to obtain high-fidelity states over broad parameter windows, while a modified Haldane–Shastry Hamiltonian with long-range interactions requires $p$9. These results are important because they show that HVA can represent both logarithmic and power-law entanglement scaling within modestly structured ansätze (Wiersema et al., 2020).

In strongly correlated fermionic systems, the 1D Hubbard-chain variational Hamiltonian ansatz alternates exponentials of on-site interaction and even/odd hopping operators,

$$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$0

Classical simulation and hardware experiments show that shallow circuits can reproduce energies and double occupancy accurately even when fidelity is moderate, but long-range spin correlations remain difficult: they may be substantially wrong even when fidelity exceeds $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$1. The same work shows that post-selection and Richardson extrapolation improve noisy estimates, especially in small systems (Martin et al., 2021).

Several newer variants broaden the HVA design space. Measurement-based HVA (MBHVA) exploits the fact that multi-qubit Pauli rotations can be implemented in MBQC with a constant number of single-qubit measurements on suitable graph states. In the 2D Heisenberg and 1D Hubbard settings, MBHVA is reported to reduce resource overhead relative to naive circuit-to-MBQC translation and to outperform a measurement-based hardware-efficient comparator (Qin et al., 2023). A commuting-group variant partitions Pauli operators into commuting clusters, diagonalizes each cluster by a Clifford unitary, and inserts parameterized single-qubit rotations in the resulting bases. In quantum chemistry benchmarks, this “combined-codes” construction reaches FCI accuracy for small molecules with one or two layers and often uses fewer two-qubit gates than a straightforward variational Hamiltonian ansatz (Vaessen et al., 2023). The truncated variational Hamiltonian ansatz (tVHA) keeps all one-body and Coulomb terms but truncates non-Coulomb two-body terms according to a cumulative-weight threshold $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$2; in the reported benchmarks, $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$3 often provides a useful accuracy–circuit-size compromise while retaining only $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$4 parameters for $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$5 Trotter steps (Possel et al., 26 May 2025).

HVA has also been adapted to excited-state and field-theoretic settings. In the quasiparticle VQE framework, an FFFT-prepared momentum-space particle-hole excitation is mapped to real space and evolved with a translation- and parity-preserving XXZ HVA having $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$6 parameters, enabling reconstruction of low-lying dispersions and velocity estimates close to Bethe-ansatz values (Velury et al., 20 Nov 2025). For $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$7d $$\ket{\psi_p(\boldsymbol{\theta})} = \prod_{l=1}^p\left(\prod_s e^{-i\theta_{s,l}H_s}\right)\ket{\psi_0}.$$8 lattice gauge theory, five HVAs built from the gauge-theory Hamiltonian respect local and global symmetries, exhibit overparameterization effects aligned with the disappearance of local minima, and show a gradient-descent loss-decay rate that scales linearly with the number of parameters (Yamanaka et al., 4 Jun 2026).

6. Assessment, misconceptions, and open problems

A recurring misconception is that HVA’s value lies only in low parameter count. The literature does not support that reductionist view. HVA’s salient property is the conjunction of Hamiltonian alignment, symmetry structure, and constrained geometry. This is why it can outperform more expressive ansätze on entanglement fidelity or trainability for some tasks, while underperforming them in others. The same restricted manifold that helps in one regime can hinder in another (Tripathi et al., 22 Apr 2026, Tripathi et al., 19 Feb 2026).

A second misconception is that accurate energy implies an accurate state. Multiple studies contradict this. In the Hubbard chain, energy and double occupancy can be accurate even when long-range spin correlations are not. In TFIM, hardware-efficient and symmetry-breaking ansätze can achieve competitive energies while misrepresenting finite-size entanglement or mixing nearly degenerate sectors. This suggests that HVA benchmarks should include energy variance, entanglement entropy, symmetry quantum numbers, and correlation functions rather than energy alone (Martin et al., 2021, Tripathi et al., 19 Feb 2026).

The main open issue is scalability under realistic optimization and hardware constraints. In 1D, HVA often works well; in 2D it becomes more fragile, and in 3D TFIM studies the authors do not use HVA directly because optimization becomes too difficult. Related work on symmetry-breaking layers shows that respecting all Hamiltonian symmetries may itself be the source of unfavorable depth scaling in symmetry-broken phases, so breaking selected symmetries can be a principled modification rather than an ad hoc one (Tripathi et al., 22 Apr 2026, Park, 2021).

Taken together, the current literature presents HVA not as a single fixed ansatz but as a design paradigm. Its core principle is stable: construct variational layers from the physically meaningful terms of the target Hamiltonian. Around that core, the field now explores symmetry-breaking layers, constrained initializations, measurement-based realizations, commuting-group restructurings, Hamiltonian truncations, and momentum-space embeddings. This suggests that the enduring significance of HVA lies less in one canonical circuit than in a family of Hamiltonian-structured ansätze whose geometry, symmetry content, and optimization behavior can be tuned to the problem class at hand (Wiersema et al., 2020, Park et al., 2023).