betaFeedback

WorksPhysics and Astronomy

Solution of the Equation-of-Motion Phonon Method eigenvalue problems on the D-Wave quantum annealer

Abstract

The solution of large-scale eigenvalue problems is crucial in nuclear many-body theory, where Hamiltonian matrices often reach extremely large dimensions. Quantum computing opens new perspectives for addressing such demanding problems. Although the Quantum Phase Estimation algorithm offers, in principle, a systematic route to matrix diagonalization, its practical deployment demands levels of coherence and error correction that current quantum hardware cannot yet support. A viable near-term strategy is instead to exploit quantum annealing, which enables the recasting of eigenvalue problems into quadratic unconstrained binary optimization formulations that can be addressed by existing annealing-based processors. Here, we propose a hybrid quantum-classical algorithm that combines quantum annealing and classical deflation to iteratively extract the full eigenspectrum of both standard and generalized eigenvalue problems. We benchmark this method on eigenvalue problems arising from the Equation of Motion Phonon Method performing calculations on real quantum hardware. Our approach illustrates the capabilities and limitations of near-term quantum devices in addressing nuclear eigenvalue problems.

INTRODUCTION

The nuclear many-body problem cannot be solved exactly for systems with more than a few nucleons [1–5], due to the extremely rapid growth of the Hilbert space dimension. As a consequence, a wide variety of many-body methods has been developed, each introducing controlled approximations to make the problem computationally tractable. The size of the Hamiltonian matrices resulting from these approaches depends both on the chosen truncation scheme and on the number of active nucleons; nevertheless, even with sophisticated approximations, the resulting matrices can reach extremely large dimensions [6]. Among others we mention the ab initio methods, such as Coupled-Cluster (CC) theory [7, 8] and the In-Medium Similarity Renormalization Group (IM-SRG) [9–11], the random-phase approximation and its extensions [12–15], the Relativistic Time-Blocking Approximation (RTBA) [16–18], and, finally, the method adopted in this work, the Equation of Motion Phonon Method (EMPM) [19].

This approach has been successfully applied to study low- and high-energy spectroscopic properties of even-even [20, 21] and odd-even nuclei [22, 23], as well as extended to the quasiparticle basis for open-shell systems [24]. The EMPM mitigates the challenges of large-scale many-body nuclear problems by constructing an orthonormal basis of multiphonon states built upon one-phonon excitations computed in the Tamm-Dancoff Approximation (TDA). By solving a set of equations of motion, EMPM captures collective correlations up to arbitrary phonon number while compressing the Hamiltonian into smaller blocks. Nevertheless, these matrices can still reach very large dimensions, and repeated diagonalizations remain computationally costly in large-scale calculations.

Classical iterative eigensolvers-such as the Lanczos [25, 26] and Davidson algorithm [27]-exploit sparsity and subspace projection to compute extremal eigenpairs without constructing the full spectrum [28]. These methods build progressively larger Krylov or Davidson subspaces and converge rapidly when spectral gaps are large. However, the storage of growing subspaces becomes a bottleneck, even with modern high-performance implementations [26, 28].

Quantum computing offers an alternative route to solve eigenvalue problems. In principle, the Quantum Phase Estimation (QPE) algorithm [29] can extract eigenvalues of Hermitian operators with arbitrary precision, provided fault-tolerant qubits and deep circuits are available. Despite its theoretical potential, QPE remains out of reach for current Noisy Intermediate-Scale Quantum (NISQ) devices due to coherence and error correction limitations [30].

To bridge the gap between quantum algorithms and today’s noisy hardware, many hybrid approaches have been developed. Most of these methods were initially designed to approximate the ground-state energy of a quantum system, with the Variational Quantum Eigensolver (VQE) and its numerous extensions playing a central role [31, 32]. Other hybrid formulations, such as the quantum version of the Lanczos algorithm [33], were likewise introduced with the primary goal of improving ground-state calculations. Building on the success of ground-state VQE, extensions aimed at computing excited states were subsequently proposed [34, 35, 36]. Despite their practical relevance, these approaches inherit several limitations of the variational framework: the optimization landscape is highly nonlinear, the energy estimators are affected by stochastic noise, and the complexity grows significantly when multiple excited states are included. Further hybrid strategies based on the Davidson algorithm have also been explored for excited-state targeting [37].

An alternative and increasingly influential direction for excited-state calculations on quantum computers is based on quantum subspace methods. This class of approaches includes quantum subspace expansion [38, 39, 40], non-orthogonal VQE [41], equation-of-motion techniques [42, 43, 44, 45], and the Quantum Krylov Subspace (QKS) framework, inspired by classical Krylov algorithms [46, 47, 48]. These methods aim to overcome some of the limitations of standard variational algorithms by projecting the problem onto a suitably constructed low-dimensional subspace, often resulting in more stable and scalable excited-state computations.

Another hybrid approach is based on quantum annealing and Quadratic Unconstrained Binary Optimization (QUBO) problems. Recent works have demonstrated the feasibility of quantum annealing for addressing both Standard (SEVPs) and Generalized EigenValue Problems (GEVPs) [49, 50, 51, 52]. In particular, the authors in [49] introduced an iterative steepest-descent approach where each step of Rayleigh quotient minimization is encoded into a QUBO instance, solved via quantum annealing. Repeated annealing steps drive convergence toward the ground-state eigenpair of a symmetric matrix, as shown in proof-of-principle demonstrations on small-scale problems [53]. However, this scheme is intrinsically limited to computing only the lowest-energy solution.

Building on this foundation, the authors in [54] demonstrated that it is possible to compute the minimal or maximal right eigenvalue-and corresponding eigenvector-of nonsymmetric GEVPs using a QUBO formulation compatible with quantum annealers. Their specific target stems from the discretization of the homogeneous Bethe-Salpeter equation (hBSE) in ladder approximation, formulated directly in four-dimensional Minkowski space. This development shows that quantum annealers, such as the D-Wave Advantage system with more than 5,000 qubits and 15-way connectivity, can be employed not only to solve symmetric GEVPs but also to tackle nonsymmetric problems relevant to relativistic bound-state equations.

In this work, we build upon the algorithm outlined in [49] and extend it to recover the full eigenspectrum for both SEVPs and GEVPs. We present a hybrid classical-quantum framework where successive eigenpairs are extracted by alternating QUBO-based quantum optimizations with classical deflation steps. After obtaining the kk-th eigenpair via quantum annealing, we project out the span of previously found eigenvectors, allowing the next QUBO call to target the subsequent eigenstate in descending order of energy magnitude.

We benchmark our extended algorithm on Hamiltonian matrices arising from EMPM calculations in nuclear structure. By varying parameters such as the harmonic-oscillator basis size and binary encoding precision, we compare simulated annealing and quantum annealing in terms of convergence behaviour, eigenvalue accuracy, and scalability. This work provides a practical strategy for exploiting near-term quantum hardware in nuclear-structure calculations. By combining classical preprocessing with quantum annealing strategies in a scalable framework, our approach opens new directions for solving large nuclear eigenvalue problems, moving closer to exploiting quantum computational resources for realistic physical applications.

The EMPM: A Brief Outline

Assuming that the (n−1)(n-1)-phonon basis states ∣α(n−1)⟩|\alpha_{(n-1)}\rangle of energies Eα(n−1)E_{\alpha_{(n-1)}} are known, we construct the set of redundant states

∣λαn−1⟩=Oλ†∣αn−1⟩,(1)|\lambda\alpha_{n-1}\rangle= O_\lambda^\dagger|\alpha_{n-1}\rangle, \tag*{(1)}

where

Oλ†=∑phcphλap†bh,(2)O_\lambda^\dagger= \sum_{ph} c^\lambda_{ph}a_p^\dagger b_h, \tag*{(2)}

creates a TD phonon of energy EλE_\lambda out of HF vacuum ∣0⟩|0\rangle through the action of the particle (ap†=axjpjmp†a_p^\dagger= a^\dagger_{x_jp_jm_p}) and hole (bh=(−)jh+mhaxhjh−mhb_h = (-)^{j_h+m_h}a_{x_hj_h-m_h}) creation operators.

We first extract from the redundant set a basis of linearly independent (though not orthogonal) states ∣λαn−1⟩|\lambda\alpha_{n-1}\rangle using the Cholesky decomposition method and use this basis to derive and solve the eigenvalue problem within the nn-subspace. To this end we start from the equations of motion

⟨αn−1∣[Oλ,H]∣αn⟩=(Eαn−Eαn−1)⟨αn−1∣αn⟩.(3)\langle\alpha_{n-1}|[O_\lambda,H]|\alpha_n\rangle= (E_{\alpha_n}-E_{\alpha_{n-1}})\langle\alpha_{n-1}|\alpha_n\rangle. \tag*{(3)}

After expanding the commutator and carrying out the standard algebraic manipulations detailed in Ref. [21], we obtain the generalized eigenvalue equations

(H−E)DC=0,(4)(\mathcal{H}-E)DC = 0, \tag*{(4)}

or more explicitly

∑jk(Hikαn−Eαnδik)DkjαnCjαn=0.(5)\sum_{jk}(\mathcal{H}^{\alpha_n}_{ik}-E_{\alpha_n}\delta_{ik})D^{\alpha_n}_{kj}C^{\alpha_n}_j = 0. \tag*{(5)}

Here

Hikαn=H(λαn−1)(λ′′αn−1′′)αn=(Eλ+Eαn−1)δλλ′′δαn−1αn−1′′+V(λαn−1)(λ′′αn−1′′)αn,(6)\mathcal{H}^{\alpha_n}_{ik}=\mathcal{H}^{\alpha_n}_{(\lambda\alpha_{n-1})(\lambda''\alpha''_{n-1})}=(E_\lambda+E_{\alpha_{n-1}})\delta_{\lambda\lambda''}\delta_{\alpha_{n-1}\alpha''_{n-1}}+\mathcal{V}^{\alpha_n}_{(\lambda\alpha_{n-1})(\lambda''\alpha''_{n-1})}, \tag*{(6)}

where V(λαn−1)(λ′′αn−1′′)αn\mathcal{V}^{\alpha_n}_{(\lambda\alpha_{n-1})(\lambda''\alpha''_{n-1})} defines the phonon-phonon interaction, and

Dijαn=D(λαn−1)(λ′αn−1′)αn=⟨λ′αn−1′∣λαn−1⟩,(7)\mathcal{D}^{\alpha_n}_{ij}=\mathcal{D}^{\alpha_n}_{(\lambda\alpha_{n-1})(\lambda'\alpha'_{n-1})}=\langle\lambda'\alpha'_{n-1}\mid\lambda\alpha_{n-1}\rangle, \tag*{(7)}

is the overlap or metric matrix which preserves the Pauli principle. The expressions of D\mathcal{D} and V\mathcal{V} can be found, for instance, in Ref. [21].

At this stage, we employ the singular value decomposition (SVD) to identify and remove spurious states [20]. The intrinsic states obtained after this projection satisfy the transformed eigenvalue equation

(H′−E)D′C′=0.(8)\left(\mathcal{H}'-E\right)D'C'=0. \tag*{(8)}

The c.m. free nn-phonon eigenstates so obtained can be recast in terms of the original basis

∣αn⟩=∑λαn−1Cλαn−1αn∣λαn−1⟩.(9)|\alpha_n\rangle=\sum_{\lambda\alpha_{n-1}}C^{\alpha_n}_{\lambda\alpha_{n-1}}|\lambda\alpha_{n-1}\rangle. \tag*{(9)}

By iterating this procedure for successive values of nn, we generate a complete orthonormal set of multiphonon state which, added to the HF vacuum (∣0⟩|0\rangle) and the TD phonons ({∣α1⟩}={∣λ⟩}\{|\alpha_1\rangle\}=\{|\lambda\rangle\}), form an orthonormal basis {∣αn⟩}\{|\alpha_n\rangle\} (n=0,1,2,3,…n=0,1,2,3,\ldots).

This basis is then used to construct and solve the eigenvalue problem in the full multiphonon space

∑αnβn′((Eαn−Eν)δαnβn′+Vαnβn′)Cβn′ν=0,(10)\sum_{\alpha_n\beta_{n'}}\left((E_{\alpha_n}-\mathcal{E}_{\nu})\delta_{\alpha_n\beta_{n'}}+\mathcal{V}_{\alpha_n\beta_{n'}}\right)C^{\nu}_{\beta_{n'}}=0, \tag*{(10)}

where Vαnβn′=0\mathcal{V}_{\alpha_n\beta_{n'}}=0 for n′=nn'=n. The coupling of the nn-phonon ∣α⟩=∣αn⟩|\alpha\rangle=|\alpha_n\rangle to the n′n'-phonon states ∣β′⟩=∣βn′⟩|\beta'\rangle=|\beta_{n'}\rangle has the structure

Vαβ′=∑λ′α′Vαα′λ′⟨λ′α′∣β′⟩,(11)\mathcal{V}_{\alpha\beta'}=\sum_{\lambda'\alpha'}\mathcal{V}^{\lambda'}_{\alpha\alpha'}\langle\lambda'\alpha'|\beta'\rangle, \tag*{(11)}

for n′=n+1n'=n+1 and

Vαβ′=∑α2⟨0∣H∣α2⟩⟨α2α∣β′⟩(12)\mathcal{V}_{\alpha\beta'}=\sum_{\alpha_2}\langle0|H|\alpha_2\rangle\langle\alpha_2\alpha|\beta'\rangle \tag*{(12)}

for n′=n+2n'=n+2. The formulas giving Vαα′λ′\mathcal{V}^{\lambda'}_{\alpha\alpha'} and ⟨0∣H∣α2⟩\langle0|H|\alpha_2\rangle, can be found elsewhere [21]. The solution of the final eigenvalue Eq. (10) yields the eigenvectors (n=0,1,2,3,…n=0,1,2,3,\ldots)

∣Ψν⟩=∑n,αnCαnν∣αn⟩.(13)|\Psi_\nu\rangle=\sum_{n,\alpha_n}C^\nu_{\alpha_n}|\alpha_n\rangle. \tag*{(13)}

QUBO-BASED ALGORITHM

Having outlined the EMPM formalism, we now describe the QUBO-based algorithm used to solve the associated eigenvalue problems. Let AA be a symmetric matrix, a well-known consequence of the spectral theorem is that the smallest eigenvalue λ\lambda and the corresponding eigenvector vv are global minima for the Rayleigh quotient ⟨x∣A∣x⟩/⟨x∣x⟩\langle x|A|x\rangle/\langle x|x\rangle, therefore

λ=min⁡∥x∥=1⟨x∣A∣x⟩,∣v⟩=argmin⁡∥x∥=1⟨x∣A∣x⟩.(14)\lambda=\min_{\lVert x\rVert=1}\langle x|A|x\rangle,\qquad|v\rangle=\underset{\lVert x\rVert=1}{\operatorname{argmin}}\langle x|A|x\rangle. \tag*{(14)}

Equivalently, one can search for the maximum magnitude of the Rayleigh quotient. Thus, for given a real symmetric matrix AA, one has to compute

λ=max⁡∥x∥=1∣⟨x∣A∣x⟩∣,∣v⟩=arg max⁡∥x∥=1∣⟨x∣A∣x⟩∣.(15)\lambda=\max_{\lVert x\rVert=1}|\langle x|A|x\rangle|,\qquad|v\rangle=\underset{\lVert x\rVert=1}{\operatorname{arg\,max}}|\langle x|A|x\rangle|. \tag*{(15)}

The maximum magnitude of a function can be found by evaluating both its minimum and maximum values and taking the one with the largest absolute value. Moreover, the maximum of a function can be conveniently computed by minimizing its negation.

To encode this minimization problem on a quantum annealer, the continuous Rayleigh quotient optimization must be reformulated in terms of binary variables to both obtain a good initial guess for the global minimum, and to implement an iterative descent from this initial estimate.

To address a real-variable optimization problem, one can encode it in the QUBO form by first approximating each real variable xx with bb binary variables. Once the number of bits bb has been set, we can build the precision vector of length bb

⟨p∣=(−1,12,122,123,…,12b−1),(16)\langle p|=\left(-1,\frac{1}{2},\frac{1}{2^2},\frac{1}{2^3},\ldots,\frac{1}{2^{b-1}}\right), \tag*{(16)}

the real variable xx can be approximated as

x=⟨p∣xb⟩,(17)x=\langle p|x_b\rangle, \tag*{(17)}

where ∣xb⟩|x_b\rangle is a binary vector of dimension bb. So, the set of all the real values we can obtain with a precision vector of dimension bb is

C1,b:={⟨p∣xb⟩∣∣xb⟩∈{0,1}b}.(18)C_{1,b}:=\{\langle p|x_b\rangle\mid|x_b\rangle\in\{0,1\}^{b}\}. \tag*{(18)}

More generally, the nn-fold product of C1,bC_{1,b} is the set

Cn,b:={(In⊗⟨p∣)∣xb⟩∣∣xb⟩∈{0,1}nb},(19)C_{n,b}:=\{(I_n\otimes\langle p|)|x_b\rangle\mid|x_b\rangle\in\{0,1\}^{nb}\}, \tag*{(19)}

where InI_n is the identity matrix and the set Cn,bC_{n,b} is a discretized cube. Now, let QQ be a symmetric n×nn\times n matrix, ∣r⟩|r\rangle an nn-vector and suppose we want to solve the constrained quadratic programming problem

argmin⁡∣x⟩∈[−1,1]n ⟨r∣x⟩+⟨x∣Q∣x⟩.(20)\underset{|x\rangle\in[-1,1]^n}{\operatorname{argmin}}\ \langle r|x\rangle+\langle x|Q|x\rangle. \tag*{(20)}

To get an approximate solution to Eq. (20) using a QUBO, we first replace the unit cube by a discretized unit cube to get the optimization problem

argmin⁡x∈Cn,b ⟨r∣x⟩+⟨x∣Q∣x⟩(21)\underset{x\in\mathbb{C}_{n,b}}{\operatorname{argmin}}\ \langle r\mid x\rangle+\langle x\mid Q\mid x\rangle \tag*{(21)}

Setting m=nbm=nb, ∣x⟩=(In⊗∣p⟩)∣xb⟩|x\rangle=(I_n\otimes|p\rangle)|x_b\rangle, and defining the precision matrix as

P=∣p⟩⟨p∣=(∣p⟩⟨p∣⟨p∣)(22)P=|p\rangle\langle p|= \begin{pmatrix} |p\rangle\\ \langle p|\\ \langle p| \end{pmatrix} \tag*{(22)}

the (21) is equivalent to the following QUBO problem

argmin⁡xb∈{0,1}m ⟨r∣(In⊗∣p⟩)∣xb⟩+⟨xb∣(Q⊗P)∣xb⟩=argmin⁡xb∈{0,1}m ⟨xb∣Diag⁡(⟨3∣In⊗∣p⟩)∣xb⟩+⟨xb∣(Q⊗P)∣xb⟩=argmin⁡xb∈{0,1}m ⟨xb∣(Diag⁡(⟨r∣In⊗∣p⟩)+Q⊗P)∣xb⟩(23)\begin{aligned} \underset{x_b\in\{0,1\}^{m}}{\operatorname{argmin}}\ \langle r|(I_n\otimes|p\rangle)|x_b\rangle+\langle x_b|(Q\otimes P)|x_b\rangle &=\\ \underset{x_b\in\{0,1\}^{m}}{\operatorname{argmin}}\ \langle x_b|\operatorname{Diag}(\langle3|I_n\otimes|p\rangle)|x_b\rangle+\langle x_b|(Q\otimes P)|x_b\rangle &=\\ \underset{x_b\in\{0,1\}^{m}}{\operatorname{argmin}}\ \langle x_b|(\operatorname{Diag}(\langle r|I_n\otimes|p\rangle)+Q\otimes P)|x_b\rangle \tag*{(23)} \end{aligned}

where Diag⁡(∣v⟩)\operatorname{Diag}(|v\rangle) refers to the diagonal matrix with entries from ∣v⟩|v\rangle. Therefore, given a number of bits bb, this procedure approximates the real optimization problem in nn variables of (20) with the QUBO of (23) of size n⋅bn\cdot b. We conclude by highlighting that we are not restricted to the cube [−1,1]n[-1,1]^n, but if we want to optimize over the cube [−δ,δ]n[-\delta,\delta]^n, we need to repeat the same construction with the precision vector δ⋅∣p⟩\delta\cdot|p\rangle.

Algorithms that solve continuous optimization problems rely on a good initial guess and an iterative descent rule. These tasks can be formulated as QUBO problems when trying to minimize the Rayleigh quotient over the unit sphere.

Initial guess phase

As in most iterative optimization procedures, a suitable initial guess is required to ensure convergence of the descent phase. Suppose that λ1≤λ2≤…≤λn\lambda_1\leq\lambda_2\leq\ldots\leq\lambda_n are the eigenvalues of AA. To get an initial approximation of λ1\lambda_1, one can ask to solve

argmin⁡x∈Cn,b ⟨x∣A∣x⟩(24)\underset{x\in\mathbb{C}_{n,b}}{\operatorname{argmin}}\ \langle x|A|x\rangle \tag*{(24)}

as approximation of

argmin⁡∥x∥=1 ⟨x∣A∣x⟩(25)\underset{\|x\|=1}{\operatorname{argmin}}\ \langle x|A|x\rangle \tag*{(25)}

To avoid the collapse of the solution to ∣x⟩=∣0⟩|x\rangle=|0\rangle when AA is positive definite we replace AA with A−λInA-\lambda I_n, where λ∈(λi,λi+1)\lambda\in(\lambda_i,\lambda_{i+1}) for some ii. This spectral shift preserves the eigenvectors, while the eigenvalues can be recovered from the new matrix

argmin⁡∣x⟩∈Cn,b ⟨x∣(A−λIn)∣x⟩(26)\underset{|x\rangle\in\mathbb{C}_{n,b}}{\operatorname{argmin}}\ \langle x|(A-\lambda I_n)|x\rangle \tag*{(26)}

A relatively good initial choice for λ\lambda is the average of the eigenvalues λ=tr⁡(A)/n\lambda=\operatorname{tr}(A)/n as suggested in [49]. This is a natural choice since it ensures that the initial guess lies within the bounds of the eigenspectrum. Nevertheless, this estimate can be refined. In fact, following the statistical analogy introduced in [55], one defines the mean mm and standard deviation ss of the spectrum as

m=Tr⁡(A)n=1n∑j=1nλj(27)m=\frac{\operatorname{Tr}(A)}{n}=\frac{1}{n}\sum_{j=1}^{n}\lambda_j \tag*{(27)}
s2=1nTr⁡(A2)−m2.(28)s^2=\frac{1}{n}\operatorname{Tr}(A^2)-m^2. \tag*{(28)}

As a consequence,

m−sn−1≤λmin⁡≤m−sn−1,m-s\sqrt{n-1}\leq\lambda_{\min}\leq m-\frac{s}{\sqrt{n-1}},
m+sn−1≤λmax⁡≤m+sn−1.(29)m+\frac{s}{\sqrt{n-1}}\leq\lambda_{\max}\leq m+s\sqrt{n-1}. \tag*{(29)}

that for n=2n=2 reduces to

λmin⁡=m−s,λmax⁡=m+s.(30)\lambda_{\min}=m-s,\qquad\lambda_{\max}=m+s. \tag*{(30)}

In general we have that,

m−sk−1n−k+1≤λk≤m+sn−kk.(31)m-s\sqrt{\frac{k-1}{n-k+1}}\leq\lambda_k\leq m+s\sqrt{\frac{n-k}{k}}. \tag*{(31)}

Once set the initial value for the eigenvalue, as the algorithm progresses, it decreases towards the target λ1\lambda_1. These observations lead to the following algorithm, based on a iterative fixed-point method. The algorithm starts with the initial guess phase, where we first solve (26) to produce a guess eigenvector ∣v1⟩|v_1\rangle. Then, we update λ\lambda using the Rayleigh quotient λ=⟨v1∣A∣v1⟩∥v1∥2\lambda=\frac{\langle v_1|A|v_1\rangle}{\|v_1\|^2} and we solve (24) again using ∣v1⟩|v_1\rangle as a bias vector to produce a second guess vector ∣v2⟩|v_2\rangle. This procedure is repeated until λ\lambda is no longer decreasing.

Iterative descent phase

After obtaining an initial approximation of the lowest eigenvalue, the algorithm refines it through an iterative descent procedure.

Suppose we want to minimize a function f:Rm→Rf:\mathbb{R}^m\to\mathbb{R}, and let ∇f\nabla f and He(f)\mathcal{H}e(f) denote the gradient and Hessian of ff, respectively. Starting with an initial guess ∣x0⟩|x_0\rangle, a common strategy is to Taylor expand ff around ∣x0⟩|x_0\rangle, namely

f(∣x⟩)=f(∣x0⟩)+∇f⋅∣δ⟩+⟨δ∣H(f)2∣δ⟩+o(∥δ∥3),(32)f(|x\rangle) = f(|x_0\rangle) + \nabla f \cdot|\delta\rangle+ \langle\delta| \frac{\mathcal{H}(f)}{2} |\delta\rangle+ o(\|\delta\|^3), \tag*{(32)}

where ∣δ⟩:=∣x⟩−∣x0⟩|\delta\rangle:= |x\rangle- |x_0\rangle, and choose a ∣δ⟩|\delta\rangle to minimize ∇f⋅∣δ⟩\nabla f \cdot|\delta\rangle (gradient descent), or ∇f⋅∣δ⟩+⟨δ∣H(f)2∣δ⟩\nabla f \cdot|\delta\rangle+ \langle\delta| \frac{\mathcal{H}(f)}{2} |\delta\rangle (second order methods such as Newton’s method, BFGS, and Newton-CG [56]). Once a better solution ∣x1⟩=∣x0⟩+∣δ⟩|x_1\rangle= |x_0\rangle+ |\delta\rangle has been found, we Taylor expand around ∣x1⟩|x_1\rangle again and repeat. The proposed algorithm obtains a good descent direction by using a QUBO to find good approximate solutions to

argmin⁡δ∈Cn,b∇f⋅∣δ⟩+⟨δ∣H(f)2∣δ⟩.(33)\underset{\delta\in C_{n,b}}{\operatorname{argmin}} \nabla f \cdot|\delta\rangle+ \langle\delta| \frac{\mathcal{H}(f)}{2} |\delta\rangle. \tag*{(33)}

Similar to Newton-CG and BFGS, this method requires computing, but not inverting, the Hessian matrix H(f)\mathcal{H}(f), and benefits from a line search which possibly increases the size of δ\delta. If the kk-th approximate solution ∣xk⟩|x_k\rangle is closer to the true solution than any point in the discretized cube, one cannot expect minimizing the QUBO to produce a better solution. If the new candidate solution ∣xk+1⟩|x_{k+1}\rangle is worse than the previous one, the algorithm discards the candidate and replaces the discretized unit cube Cn,bC_{n,b} by a scaled-down discretized cube t⋅Cn,bt \cdot C_{n,b}, where t≪1t \ll1, which amounts to repeating the procedure outlined above with the new precision vector t∣p⟩t|p\rangle. Once the discretized cube has been scaled down, the algorithm continues running until it needs to scale down the cube further, or terminates once the desired accuracy is reached.

Extension to the generalized eigenvalue problem

In many cases, one needs to solve a generalized eigenvalue problem of the form A∣v⟩=λB∣v⟩A|v\rangle= \lambda B|v\rangle, where AA, BB are symmetric and BB is Strictly Positive Definite (SPD). Under these conditions, the smallest generalized eigenvalue minimizes the generalized Rayleigh quotient

λ=min⁡∥x∥=1⟨x∣A∣x⟩⟨x∣B∣x⟩,∣v⟩=argmin⁡∥x∥=1⟨x∣A∣x⟩⟨x∣B∣x⟩.(34)\lambda= \min_{\|x\|=1} \frac{\langle x|A|x\rangle}{\langle x|B|x\rangle}, \qquad|v\rangle= \operatorname{argmin}_{\|x\|=1} \frac{\langle x|A|x\rangle}{\langle x|B|x\rangle}. \tag*{(34)}

Therefore, the following small changes adapt the algorithm discussed so far to solve the generalized eigenvalue problem with essentially arbitrary precision. First, instead of m=tr⁡(A)/nm = \operatorname{tr}(A)/n, one can generate a random unit vector ∣w⟩|w\rangle and compute mm as m=⟨w∣A∣w⟩⟨w∣B∣w⟩m = \frac{\langle w|A|w\rangle}{\langle w|B|w\rangle}. Second, we have to replace every Rayleigh quotient with the corresponding generalized Rayleigh quotient. Lastly, instead of updating HH as H=A−λInH = A - \lambda I_n, we update H=A−λBH = A - \lambda B, since this preserves the BB-eigenspectrum of AA, whereas the former does not.

Matrix deflation techniques

The previously discussed algorithm computes a single eigenpair. To extend this approach to the computation of the full eigenspectrum, we employ matrix deflation techniques, which enable the sequential extraction of eigenvalues and eigenvectors.

Deflation is a classical strategy in numerical linear algebra: once an eigenpair has been computed, the system matrix is modified so that the same eigenpair does not reappear in subsequent iterations. This enables iterative procedures—such as the QUBO-based eigenvalue search described earlier—to identify additional eigenpairs one at a time.

Our use of deflation is motivated by the structure of the QUBO formulation. Unlike global eigensolvers (e.g., the QZ [57, 58] algorithm or dense QR methods [59, 60]), which compute multiple eigenpairs simultaneously, the QUBO solver isolates a single eigenpair per optimization instance. After each eigenpair is found, deflation ”peels away” its influence from the matrix, allowing the algorithm to target the next dominant mode without interference.

Although a fully quantum deflation framework could, in principle, be implemented as part of a future quantum eigensolver, such an extension lies beyond the scope of this work.

Deflation techniques can be broadly grouped into two main categories:

  • Subtractive deflation: The contribution of the computed eigenpair is subtracted from the matrix to eliminate its influence. Examples include Hotelling’s deflation and Orthogonal Projection deflation.

  • Orthogonalization-based deflation: The search space for subsequent eigenvector estimates is restricted to vectors orthogonal to the previously computed eigenvectors. Householder deflation is a representative method in this class.

In our procedure, we implement three classical and widely-used deflation techniques—Hotelling, Orthogonal Projection, and Householder deflation—covering both categories above. These methods are employed for both the standard and the generalized eigenvalue problems, enabling the sequential reconstruction of the full eigenspectrum as detailed in § 6. A summary of their main properties is provided in Table I.

Hotelling Deflation

Hotelling deflation is a technique used in eigenvalue problems to find additional eigenvalues after the dominant eigenpair (eigenvalue λ1\lambda_1 and eigenvector ∣v1⟩|v_1\rangle) has been found. It modifies the original matrix to ”remove” the influence of the found eigenpair. Given a matrix AA and its dominant eigenpair (λ1,∣v1⟩)(\lambda_1, |v_1\rangle), with ∣v1⟩|v_1\rangle normalized so that ∥v1∥=1\lVert v_1\rVert= 1, the deflated matrix A′A' is defined as

FeatureHotellingOrthogonal ProjectionHouseholder
Basic IdeaSubtracts the rank-1 update formed by the found eigenpair from the matrixProjects the matrix onto the subspace orthogonal to found eigenvectorsUses Householder reflections to transform the matrix, removing found eigenvectors
TypeSubtractiveSubtractiveOrthogonalization-based
Mathematical Operation for SEVPA′=A−λ1∣v1⟩⟨v1∣A' = A-\lambda_1 |v_1\rangle\langle v_1|, where λ\lambda, ∣v1⟩|v_1\rangle are eigenvalue/vectorA′=(I−∣v1⟩⟨v1∣)A(I−∣v1⟩⟨v1∣)A' = (I-|v_1\rangle\langle v_1|) A (I-|v_1\rangle\langle v_1|), where ∣v1⟩|v_1\rangle is the eigenvectorApplies orthogonal Householder transformation HH to deflate, where H=I−2∣x⟩⟨x∣∥x∥2H = I-2\frac{|x\rangle\langle x|}{\lVert x\rVert^2}, ∣x⟩=∣v⟩±∥v∥2∣e1⟩|x\rangle= |v\rangle\pm\lVert v\rVert_2 |e_1\rangle and ∣v⟩|v\rangle is the eigenvector
Mathematical Operation for GEVPA′=A−λ1B∣v1⟩⟨v1∣BTA' = A-\lambda_1 B |v_1\rangle\langle v_1| B^T, where λ\lambda, ∣v1⟩|v_1\rangle are eigenvalue/vectorA′=(I−B∣v1⟩⟨v1∣)A(I−B∣v1⟩⟨v1∣)A' = (I-B|v_1\rangle\langle v_1|) A (I-B|v_1\rangle\langle v_1|), B′=(I−A∣v1⟩⟨v1∣)B(I−A∣v1⟩⟨v1∣)B' = (I-A|v_1\rangle\langle v_1|) B (I-A|v_1\rangle\langle v_1|), where ∣v1⟩|v_1\rangle is the eigenvectorApplies orthogonal Householder transformation HH to both AA and BB matrices as follows: A←HAHA \leftarrow H A H, B←HBHB \leftarrow H B H
Preserves Symmetry?YesYesYes
Numerical StabilityModerate; can introduce errors due to subtractionHigher stability due to projectionHigh stability from orthogonal transformations
Ease of ImplementationSimpleMore complex; requires projection operationsMore complex; requires constructing Householder reflectors

Table 1. Comparison of Hotelling, Orthogonal Projection, and Householder deflation methods.

A′=A−λ1∣v1⟩⟨v1∣.(35)A' = A-\lambda_1 |v_1\rangle\langle v_1| . \tag*{(35)}

As result, the deflated matrix A′A' will have ∣v1⟩|v_1\rangle as an eigenvector corresponding to an eigenvalue of zero. All other eigenvalues of AA (and their corresponding eigenvectors) remain unchanged in A′A'. It is conceptually straightforward, but stability is its major weakness. Hotelling deflation is highly sensitive to errors in the computed λ1\lambda_1 and ∣v1⟩|v_1\rangle. Small inaccuracies can accumulate rapidly, leading to poor accuracy for subsequently found eigenpairs. This makes it less suitable for accurately finding many eigenvalues.

To extend this deflation technique to the GEVP, the matrix BB must be introduced. A common form of Hotelling-like deflation often involves modifying AA while keeping BB the same. If we have A∣v1⟩=λ1B∣v1⟩A|v_1\rangle= \lambda_1 B|v_1\rangle, with ∣v1⟩|v_1\rangle normalized such that ⟨v1∣B∣v1⟩=1\langle v_1|B|v_1\rangle= 1, a Hotelling-like deflated matrix A′A' can be defined as

A′=A−λ1B∣v1⟩⟨v1∣BT.(36)A' = A-\lambda_1 B|v_1\rangle\langle v_1|B^T. \tag*{(36)}

Numerical instability, as in the standard problem, remains the primary issue, because the subtraction of terms like λ1B∣v1⟩⟨v1∣BT\lambda_1 B|v_1\rangle\langle v_1|B^T assumes perfect knowledge of λ1\lambda_1 and ∣v1⟩|v_1\rangle. In practice, computed eigenpairs are only approximate. Errors in λ1\lambda_1 and ∣v1⟩|v_1\rangle are introduced at each deflation step, and these errors can accumulate rapidly.

Orthogonal Projection Deflation

Orthogonal Projection deflation is a technique used in SEVP to find multiple eigenpairs by ensuring that subsequent eigenvector estimates are orthogonal to previously found eigenvectors. This is done by projecting the matrix onto the subspace orthogonal to the space spanned by the known eigenvectors.

Given a matrix AA and its dominant eigenpair (λ1,∣v1⟩)(\lambda_1, |v_1\rangle), with ∣v1⟩|v_1\rangle normalized so that ∥v1∥=1\lVert v_1\rVert= 1, the deflated matrix A′A' is defined as

A′=(I−∣v1⟩⟨v1∣)A(I−∣v1⟩⟨v1∣),(37)A' = (I-|v_1\rangle\langle v_1|) A (I-|v_1\rangle\langle v_1|), \tag*{(37)}

where II is the identity matrix.

This projection removes the influence of the known eigenvectors, allowing iterative methods to converge to new eigenpairs orthogonal to those already found. While conceptually useful and offering some advantages over the most basic Hotelling deflation, it is generally less numerically stable than Householder deflation and is not the primary method employed by state-of-the-art numerical libraries for computing complete eigenspectra (dense matrices) or subsets (sparse matrices).

For GEVP, Orthogonal Projection deflation removes the influence of already computed eigenvectors by projecting onto the subspace orthogonal to them with respect to the BB-inner product. This is achieved by defining the projection operators

Pb=I−B∣v1⟩⟨v1∣,Pa=I−A∣v1⟩⟨v1∣,(38)P_b = I-B|v_1\rangle\langle v_1|,\qquad P_a = I-A|v_1\rangle\langle v_1|, \tag*{(38)}

and applying them to the matrices AA and BB

A′=PbAPb,B′=PaBPa.(39)A' = P_b A P_b,\qquad B' = P_a B P_a. \tag*{(39)}

Householder Deflation

Householder deflation is another technique used in the SEVP to find additional eigenvalues, but employing Householder transformations. These transformations are orthogonal reflections that zero out components of vectors, allowing the matrix to be iteratively reduced in a numerically stable manner while preserving symmetry and orthogonality. By applying Householder deflation, the influence of already computed eigenvectors is effectively removed, which facilitates the computation of subsequent eigenpairs. Given a matrix AA and an eigenvector v1v_1, a Householder reflector HH is constructed such that

H∣v1⟩=α∣e1⟩,(40)H\lvert v_1\rangle= \alpha\lvert e_1\rangle, \tag*{(40)}

where ∣e1⟩\lvert e_1\rangle is the first canonical basis vector and α=±∥v1∥2\alpha= \pm\lVert v_1\rVert_2. Applying HH to AA produces a transformed matrix

G=HAHT.(41)G = HAH^{T}. \tag*{(41)}

The transformed matrix GG will have a block structure where λ1\lambda_1 is isolated, and the remaining (n−1)×(n−1)(n-1)\times(n-1) submatrix A′A' contains the other eigenvalues of AA. Iterative deflation steps are then applied to A′A'.

Householder deflation is numerically stable due to the orthogonality of HH and is widely used in QR algorithms and advanced eigensolvers for SEVP.

For GEVP we apply Householder reflectors to both AA and BB matrices as follows

A←HAH,B←HBH.(42)A \gets HAH,\qquad B \gets HBH. \tag*{(42)}

However, as can be seen later in § 7 B, this approach distorts the original symmetry. Applying a single Householder transformation simultaneously to both AA and BB does not directly achieve deflation of an eigenpair in a clean and isolated manner. More specifically, such a transformation does not preserve the essential structural properties of the GEVP, such as maintaining BB as a symmetric positive definite metric, or reducing AA and BB to triangular form. The objective of deflation is not only to eliminate the contribution of a specific eigenpair, but to do so while preserving the generalized eigenvalue structure and the mathematical relationship between AA and BB. For this reason, modern approaches such as the generalized Schur decomposition (QZ algorithm) are preferred for the systematic computation of the eigenspectrum in the generalized case.

Hybrid Algorithm for Full Spectrum Eigenpair Computation

In this section, we present a hybrid quantum-classical eigensolver algorithm designed to compute the full spectrum of eigenvalues and corresponding eigenvectors for both SEVP and GEVP. The proposed approach combines the algorithm presented in Section III with iterative eigenpair extraction via matrix deflation, using one of the deflation techniques described in § 5. The algorithm is termed hybrid as it combines a classical numerical procedure, matrix deflation, with a quantum-based formulation, namely the QUBO formulation of the Rayleigh quotient optimization problem.

Algorithm 1 Dominant eigenpair finder for the SEVP

Require: Symmetric matrix AA

Ensure: Eigenvalue λmm\lambda_{\mathrm{mm}} with maximum magnitude and corresponding eigenvector ∣vmm⟩\lvert v_{\mathrm{mm}}\rangle

  1. Solve QUBO to maximize the Rayleigh quotient:

∣vmax⁡⟩=arg max⁡∥x∥=1 ⟨x∣A∣x⟩,λmax⁡=⟨vmax⁡∣A∣vmax⁡⟩⟨vmax⁡∣vmax⁡⟩\lvert v_{\max}\rangle= \underset{\lVert x\rVert=1}{\operatorname{arg\,max}}\ \langle x\lvert A\lvert x\rangle,\qquad \lambda_{\max} = \frac{\langle v_{\max}\lvert A\lvert v_{\max}\rangle}{\langle v_{\max}\lvert v_{\max}\rangle}
  1. Solve QUBO to minimize the Rayleigh quotient:

∣vmin⁡⟩=arg min⁡∥x∥=1 ⟨x∣A∣x⟩,λmin⁡=⟨vmin⁡∣A∣vmin⁡⟩⟨vmin⁡∣vmin⁡⟩\lvert v_{\min}\rangle= \underset{\lVert x\rVert=1}{\operatorname{arg\,min}}\ \langle x\lvert A\lvert x\rangle,\qquad \lambda_{\min} = \frac{\langle v_{\min}\lvert A\lvert v_{\min}\rangle}{\langle v_{\min}\lvert v_{\min}\rangle}
  1. if ∣λmax⁡∣≥∣λmin⁡∣\lvert\lambda_{\max}\rvert\geq\lvert\lambda_{\min}\rvert then

  2. λmm←λmax⁡\lambda_{\mathrm{mm}} \gets\lambda_{\max}, ∣vmm⟩←∣vmax⁡⟩\lvert v_{\mathrm{mm}}\rangle\gets\lvert v_{\max}\rangle

  3. else

  4. λmm←λmin⁡\lambda_{\mathrm{mm}} \gets\lambda_{\min}, ∣vmm⟩←∣vmin⁡⟩\lvert v_{\mathrm{mm}}\rangle\gets\lvert v_{\min}\rangle

  5. end if

  6. return λmm,∣vmm⟩\lambda_{\mathrm{mm}},\lvert v_{\mathrm{mm}}\rangle

Algorithm 2 Hybrid full spectrum solver for the SEVP

Require: Symmetric matrix AA, and a chosen deflation method {Hotelling, Orthogonal Projection, Householder}\{\text{Hotelling, Orthogonal Projection, Householder}\}

Ensure: Eigenvalues {λi}\{\lambda_i\} and eigenvectors {∣vi⟩}\{\lvert v_i\rangle\}

  1. Initialize set of computed eigenpairs: S←∅\mathcal{S}\gets\varnothing

  2. Initialize deflated matrices: (A0)←(A)(A_0)\gets(A)

  3. for i=1i=1 to desired number of eigenpairs do

  4. Find eigenpair (λi,∣vi⟩)(\lambda_i,\lvert v_i\rangle) using Algorithm 1:

λi=max⁡∥x∥=1⟨x∣A∣x⟩,∣vi⟩=arg max⁡∥x∥=1⟨x∣A∣x⟩\lambda_i=\underset{\lVert x\rVert=1}{\max}\langle x\lvert A\lvert x\rangle,\qquad \lvert v_i\rangle=\underset{\lVert x\rVert=1}{\operatorname{arg\,max}}\langle x\lvert A\lvert x\rangle
  1. Add eigenpair to the set: S←S∪{(λi,∣vi⟩)}\mathcal{S}\gets\mathcal{S}\cup\{(\lambda_i,\lvert v_i\rangle)\}

  2. Apply deflation to (Ai−1)(A_{i-1}) to obtain (Ai)(A_i)

  3. end for

  4. return S\mathcal{S}

While in orthogonalization-based deflation (like Householder deflation), finding the minimum of the Rayleigh quotient at each iteration is sufficient to proceed, since the corresponding eigenpair is fully removed from the problem, this is not the case for subtractive deflation methods. In the latter, each iteration annihilates (but does not fully eliminate) the influence of the computed eigenpair. As a result, the same eigenpair might reappear as the minimum in subsequent iterations. If the matrix AA is positive definite, this issue typically manifests already at the first deflation step. To overcome this issue, we employed the QUBO-based Algorithm 1, to search for the maximum magnitude of the Rayleigh quotient, which corresponds to identifying the dominant eigenpair. Taking into account all the considerations discussed above, we formulate Algorithm 2 and Algorithm 3, which are designed to compute the complete eigenspectrum of the SEVP and the GEVP, respectively.

A potential performance issue arises because computing the maximum magnitude requires solving two QUBO problems at each iteration, resulting in a total of 2×N2\times N

Algorithm 3 Hybrid full spectrum solver for the GEVP

Require: Symmetric matrix pair (A,B)(A,B), where BB is SPD, and a chosen deflation method ∈{Hotelling,Orthogonal Projection}\in\{\text{Hotelling},\text{Orthogonal Projection}\}

Ensure: Eigenvalues {λi}\{\lambda_i\} and eigenvectors {∣vi⟩}\{|v_i\rangle\}

  1. Initialize set of computed eigenpairs: S←∅S\leftarrow\varnothing

  2. Initialize deflated matrices: (A0,B0)←(A,B)(A_0,B_0)\leftarrow(A,B)

  3. for i=1i=1 to desired number of eigenpairs do

  4. Find eigenpair (λi,∣vi⟩)(\lambda_i,|v_i\rangle) using Algorithm 1, extended for the GEVP:

λi=max⁡∥x∥=1⟨x∣A∣x⟩⟨x∣D∣x⟩,∣vi⟩=arg⁡max⁡∥x∥=1⟨x∣A∣x⟩⟨x∣D∣x⟩\lambda_i=\max_{\lVert x\rVert=1}\frac{\langle x|A|x\rangle}{\langle x|D|x\rangle},\qquad|v_i\rangle=\arg\max_{\lVert x\rVert=1}\frac{\langle x|A|x\rangle}{\langle x|D|x\rangle}
  1. Add eigenpair to the set: S←S∪{(λi,∣vi⟩)}S\leftarrow S\cup\{(\lambda_i,|v_i\rangle)\}

  2. Apply deflation to (Ai−1,Bi−1)(A_{i-1},B_{i-1}) to obtain (Ai,Bi)(A_i,B_i)

  3. end for

  4. return SS

QUBO problems for NN iterations. To mitigate this computational cost, we implement a strategy where, at each iteration, the QUBO result that is not immediately used (i.e., the eigenpair corresponding to the minimum magnitude) is stored and reused in the subsequent iteration. This approach reduces the total number of QUBO problems to solve, thereby improving computational efficiency.

Specifically, for the SEVP, only NN iterations are required if the matrix AA is strictly positive definite or strictly negative definite, whereas N+1N+1 iterations are needed if AA is not strictly positive or negative definite. The same applies to the GEVP: only NN iterations are needed if the matrix pair (A,B)(A,B) satisfies strict definiteness conditions; otherwise, N+1N+1 iterations are necessary.

Hereafter, we analyse the computational cost of each of the aforementioned deflation methods. All three methods share the same asymptotic complexity when extracting the full spectrum of dense matrices:

  • Total time: O(n3)O(n^3) for each method (arising from O(n2)O(n^2) work per deflation step, repeated ≃n\simeq n times; this excludes the cost of the eigensolver used to compute individual (λ,v)(\lambda,v) pairs).

  • Memory: O(n2)O(n^2) (storage of AA, BB, and dense aux-

iliary matrices).

Despite sharing the same asymptotic cost, the methods differ in numerical robustness, hidden constants, and structure preservation.

The Hotelling method computes u=Bvu = Bv and performs a rank-one correction of AA at each deflation step, for a cost of O(n2)O(n^2). It is simple and has low hidden constants, but may introduce numerical instabilities unless combined with re-orthogonalization or filtering.

The Orthogonal Projection approach also performs u=Bvu = Bv, followed by orthogonalization in the BB-inner product and projection, again yielding an O(n2)O(n^2) cost per step. Its cost is stable across iterations, and it is typically the most numerically robust method, producing fewer spurious eigenpairs.

Finally, the Householder deflation applies congruent reflectors to progressively smaller active submatrices. The ii-th iteration costs O(m⋅n)O(m \cdot n) with m=n−im = n - i, leading to a total of O(n3)O(n^3) but with smaller leading constants because later steps are cheaper.

This approach preserves symmetry and congruence and is often efficient in practice on dense problems, though repeated updates of AA and BB may compromise the orthogonality of the computed eigenvectors.

Numerical Implementation and Results

As a study case, we consider 4He{}^{4}\mathrm{He}. The starting Hamiltonian is defined as

H=Tint+VD.(43)H = T_{\mathrm{int}} + V_D . \tag*{(43)}

Here, TintT_{\mathrm{int}} denotes the intrinsic kinetic energy, and VDV_D is the Daejeon interaction derived from the NNNN component of the N3LO potential [61] through a two-step procedure. First, the NNNN potential is softened using the SRG method [62] with flow parameter λ=1.5 fm−1\lambda= 1.5\ \mathrm{fm}^{-1} then, a phase-equivalent transformation is applied to determine an optimal parametrization of the NNNN force. We use this Hamiltonian to generate a HF basis encompassing 3 major shells. We use the HF states to create the TD phonon basis and generate the 2-phonon basis by deriving and solving iteratively the EMPM Eq. (5); the basis so constructed is adopted to solve the final eigenvalue problem in the multiphonon space Eq. (10).

Ground state

We begin by evaluating the performance of the QUBO-based algorithm for computing the ground-state eigenpair using a D-Wave quantum annealer. All quantum annealing (QA) calculations were carried out on the Advantage System 6.4 [63], which provides 5,614 working qubits. For comparison, we also performed the same simulations using a classical Simulated Annealing (SA) solver.

The analysis covers different multipolarities, corresponding to different matrix dimensions dd, and two bit resolutions, b=2b = 2 and b=4b = 4. The starting point in the plots corresponds to the output of the initial-guess phase, where the precision parameter pp (see Section III) is set to 1. The subsequent evolution shows how the accuracy improves during the descent phase as the discretized search cube is progressively refined by increasing pp.

For each configuration, we monitor the errors on the eigenvalue (eval_error) and eigenvector (evec_error) as functions of both pp and of the iteration number (NitN_{it}). Here, eval_error denotes the deviation of the computed eigenvalue from the exact one, while evec_error is the Euclidean distance between the computed eigenvector and the true ground-state eigenvector. Reference solutions were obtained using standard numerical eigensolvers.

We first discuss the results shown in Figure 1 for b=2b = 2. The figure reports the number of correct digits of the lowest eigenvalue and the corresponding eigenvector for the TDA Hamiltonian, comparing SA and QA. Panels (a), (c), and (e) refer to SA, while panels (b), (d), and (f) refer to QA. The parameter ϵp=−log⁡10(precision)\epsilon_p = -\log_{10}(\text{precision}) is also indicated. As test cases, we consider three Hamiltonians: the TDA Hamiltonian for Jπ=0+J^\pi= 0^+ with dimension d=4d = 4, the TDA Hamiltonian for Jπ=1−J^\pi= 1^- with dimension d=7d = 7, and the full Hamiltonian for Jπ=0+J^\pi= 0^+ with dimension d=25d = 25.

Digits of accuracy of the lowest eigenvalue and corresponding eigenvector for the TDA Hamiltonian diagonalization

Figure 1. Digits of accuracy of the lowest eigenvalue and of the corresponding eigenvector for the TDA Hamiltonian diagonalization with Jπ=0+J^{\pi}=0^{+} and b=2b=2, performed with SA (a,c,e) and with QA (b,d,f). The value εp=−log⁡10(precision)\varepsilon_p=-\log_{10}(\mathrm{precision}) is also reported.

Let us start with Jπ=0+J^\pi= 0^+ (panels (a),(b)). We observe no significant difference between SA and QA in terms of the number of iterations (Nit≃14N_{it} \simeq14).

As the dimension increases with Jπ=1−J^\pi= 1^- (panels (c),(d)), QA (panel (d)) begins to show better performance compared to SA (panel (c)), with the number of iterations decreasing from Nit≃40N_{it} \simeq40 with SA to Nit≃12N_{it} \simeq12 with QA.

The most interesting scenario is the full Hamiltonian case (panels (e), (f)). Here, SA (panel (e)) fails to improve the accuracy of the lowest energy eigenvalue beyond the initial guess phase. On the other hand, QA (panel (f)) achieves machine-precision accuracy on the computed eigenvalue within Nit≃30N_{it} \simeq30 iterations.

We now analyse the case with increased bit resolution, setting b=4b = 4 (Figure 2). Compared to the b=2b = 2 case, we observe a significant improvement in the performance of QA over SA. For the largest matrix size d=25d = 25, SA (panel (e)) requires approximately Nit≃200N_{it} \simeq200 iterations to achieve machine-precision accuracy on the lowest eigenvalue, whereas QA (panel (f)) reaches the same level of accuracy within only Nit≃30N_{it} \simeq30 iterations.

Six-panel plots of digits of accuracy versus iterations for $b=4$

Figure 2. Same as Figure 1, for b=4b = 4.

Let’s conclude this Section with an analysis of QA performance in solving a GEVP. Specifically, we discuss the results obtained for the EMPM Hamiltonian with N=2N = 2 and Jπ=0+,1−J^{\pi} = 0^+, 1^- and 2+2^+ of dimension d=22,34,48d = 22, 34, 48, respectively.

Even in the GEVP case, we find a meaningful improvement in performance when using QA over SA. This is already clear for b=2b = 2 (Figure 3), where using QA we get the eigenvalue with accuracy 10−810^{-8} in Nit≃30N_{it} \simeq30, while with SA we need Nit≃450N_{it} \simeq450, as shown in Figure 3(a),(Figure 3b) for Jπ=0+J^{\pi} = 0^+. For the eigenvector, the level of accuracy is 10−710^{-7} for Nit≃60N_{it} \simeq60 with QA, versus 10−710^{-7} for Nit≃500N_{it} \simeq500 with SA. For Jπ=1−J^{\pi} = 1^- (Figure 3), we can formulate similar observations when comparing the result for SA, in Figure 3(c), with the one for QA in Figure 3(d). The SA is not sufficient to achieve 8-digit accuracy, while using QA the algorithm is able to gain 10−810^{-8} in accuracy for Nit≃40N_{it} \simeq40. An analogous behaviour comes out also for Jπ=2+J^{\pi} = 2^+, shown in Figure 3(e) for SA and in Figure 3(f) for QA. The error on the lowest eigenvector reflects the trend described for the eigenvalue error and it remains above 10−710^{-7} in all the cases.

Results for the solution of the GEVP within the Equation-of-Motion Phonon Method

Figure 3. Results for the solution of the GEVP within the Equation-of-Motion Phonon Method with Jπ=0+,1−,2+J^\pi= 0^+,1^-,2^+ and b=2b = 2.

Our results demonstrate a significant performance advantage of QA algorithm over SA also when solving GEVPs with b=4b = 4, as shown in Figure 4. Here, QA achieves an eigenvalue accuracy of 10−810^{-8} within only Nit≃20N_{it} \simeq20, while SA requires Nit≃90N_{it} \simeq90 (Figure 4(a) and (Figure 4b) for Jπ=0+J^{\pi} = 0^+). Similarly, QA finds the eigenvector with an accuracy of 10−710^{-7} for Nit≃60N_{it} \simeq60 compared to Nit≃230N_{it} \simeq230 needed by SA to reach the same level of accuracy.

Six-panel plots of digits of accuracy versus iterations for $d=22.0$, $34.0$, and $48.0$, with $b=4.0$

Figure 4. Same as Figure 3, for b=4b = 4.

Similar observations hold for Jπ=1−J^{\pi} = 1^- and Jπ=2+J^{\pi} = 2^+ (Figure 4(c-Figure 4f)). While SA struggles to achieve 8-digit accuracy in these cases, requiring from Nit≃350N_{it} \simeq350 to Nit≃1200N_{it} \simeq1200, QA readily converges to an accuracy of 10−810^{-8} within Nit≃20N_{it} \simeq20 to Nit≃30N_{it} \simeq30.

From the above analysis, we can conclude that in all examined cases, the use of QA over SA guarantees a quicker convergence to the 10−810^{-8} accuracy, since it requires less iterations. This gap in performance becomes more noticeable as the matrix size increases. In particular, there are cases in which SA doesn’t even let reach the eigenvalue accuracy of 10−810^{-8}, as shown in Figure 4(c),(Figure 4e) for Jπ=1−,2+J^{\pi}=1^{-},2^{+}. In all the other cases, the boost in performances found ranges between 78% and 98% for the number of iterations, indicating a clear improvement in sampling efficiency of QA with respect to SA.

Full spectrum

We now analyse the results for the full spectrum. Due to the limited Quantum Processing Unit (QPU) time available on the D-Wave quantum annealer, in this section we only demonstrate the reliability of our approach for the SA case. Here, the QUBO algorithm used to determine the dominant eigenpair (Algorithm 1) is combined with the three different deflation methods outlined in § 5.

We start by analysing the results obtained for the SEVP. Here, we report the results for the Jπ=0+J^{\pi}=0^{+} EMPM full Hamiltonian with b=2b=2 (Figure 5) and b=4b=4 (Figure 6). A detailed inspection of the results shows that all the implemented deflation techniques yield highly accurate final eigenvalues and eigenvectors, achieving up to 13 and 8 correct digits, respectively.

Digits of accuracy of the eigenpairs for the $J^{\pi}=0^{+}$ SEVP EMPM full Hamiltonian with respect to the three different methods adopted, with $b=2$

Figure 5. Digits of accuracy of the eigenpairs of the Jπ=0+J^{\pi}=0^{+} SEVP EMPM full Hamiltonian with respect to the three different methods adopted, with b=2b=2.

Digits of accuracy of the eigenpairs for the $J^{\pi}=0^{+}$ SEVP EMPM full Hamiltonian with respect to the three different methods adopted, with $b=4$

Figure 6. Same as Fig. 5, for b=4b=4.

The total execution time (tett_{et}), comprising both the deflation step and the annealing phase time, varies across the three deflation methods. For the Hotelling method,

tet=167.13t_{et}=167.13 s with b=2b=2 and it increases to 526.06 s with b=4b=4. For the Orthogonal Projection method, tet=155.47t_{et}=155.47 s with b=2b=2 and it increases to 533.20 s with b=4b=4. Lastly, the Householder method is significantly faster, with tet=85.53t_{et}=85.53 s with b=2b=2 and tet=211.39t_{et}=211.39 s with b=4b=4. This performance advantage stems from its deflation strategy, which progressively reduces the size of the active submatrix, thereby minimizing computational effort as the algorithm advances.

In the case of the GEVP, at variance with the SEVP, the repeated application of Householder reflectors for deflation introduces a progressive deterioration of the matrix BB that compromises numerical stability. Although each reflector

H=I−2∣u⟩⟨u∣⟨u∣u⟩,(44)H = I - 2 \frac{|u\rangle\langle u|}{\langle u \mid u\rangle}, \tag*{(44)}

is orthogonal, the transformations, reported in Eq. (42), that it generates, induce fill-in and distort the original symmetry (see, e.g., [64, 65]). Furthermore, rounding errors introduced in the early deflations are amplified in the subsequent ones, preventing already deflated vectors from remaining normalized with respect to the BB-inner product. Additionally, when eigenvalues lie close together [66], the procedure tends to fail to converge to the target eigenvalue, since adjacent spectral values accentuate small perturbations in the basis and make it difficult to correctly isolate the desired eigenvalue. As a result, the Householder deflation procedure is unstable and not applicable in this phase of the numerical comparison.

The other two methods, Hotelling and Orthogonal Projection, instead, produce results with high accuracy, comparable to the standard case.

Digits of accuracy of the eigenpairs of the GEVP EMPM with $b=4$

Figure 8. Same as Figure 7, for b=4b = 4.

Digits of accuracy of the eigenpairs of the GEVP EMPM with $J^\pi = 0^+, 1^-, 2^+$ for $b=2$

Figure 7. Digits of accuracy of the eigenpairs of the GEVP EMPM with Jπ=0+,1−,2+J^\pi= 0^+, 1^-, 2^+ with respect to the three different methods adopted, with b=2b = 2.

The total execution time (tett_{et}) for the three deflation methods varies as follows. For the Hotelling method, tet=963.11t_{et} = 963.11 s with b=2b = 2 and it increases to 1293.641293.64 s with b=4b = 4. For the Orthogonal Projection method, tet=914.56t_{et} = 914.56 s with b=2b = 2 and it increases to 1024.781024.78 s with b=4b = 4.

b = 4. In contrast, the Householder method remains significantly faster, with tet=181.22t_{et} = 181.22 s with b=2b = 2 and tet=443.97t_{et} = 443.97 s with b=4b = 4.

Conclusion

In this manuscript we have applied a QUBO-based hybrid quantum-classical framework for computing the eigenspectrum of realistic EMPM Hamiltonians. The first part of the analysis focused on the ground-state estimation, where direct comparison between simulated annealing and quantum annealing reveals that the quantum approach yields systematically better results. This clearly highlights an advantage of quantum sampling with respect to classical simulated annealing.

Building upon the validated ground-state methodology, we have then extended the approach to full-spectrum reconstruction via an iterative deflation strategy. Our findings demonstrate that the proposed QUBO-based iterative scheme, combined with classical postprocessing, accurately recovers all eigenpairs of the EMPM Hamiltonian.

In the SEVP case, all three deflation strategies-Hotelling, Orthogonal Projection and Householder-yield eigenvalues and eigenvectors with machine precision for both b=2b = 2 and b=4b = 4, confirming the numerical robustness of the method for the solution of symmetric eigenvalue problems. In contrast, for the GEVP, the progressive loss of orthogonality induced by repeated Householder updates leads to rapid numerical deterioration. Consequently, Householder deflation cannot be reliably employed in this context. The Hotelling and Orthogonal Projection methods, instead, preserve accuracy across all iterations and remain effective also in the generalized case.

Overall, these findings confirm that the proposed hybrid QUBO-deflation framework provides a viable route to extend quantum annealing methods beyond ground-state estimation toward systematic reconstruction of the entire spectrum. The comparison between simulated annealing and quantum annealing indicates a performance advantage of the quantum approach in exploring the energy landscape associated with the QUBO formulation of the EMPM Hamiltonian. Although this result is obtained within the present QUBO framework and for matrix sizes compatible with current quantum annealing hardware, it supports the potential of quantum annealing as an effective strategy for spectral estimation problems of this class.

The demonstrated robustness of the deflation mechanisms establishes the required groundwork for future implementations fully executed on quantum annealers. In particular, a natural next step will be to reduce the degree of classical part of the algorithm by encoding also the deflation step directly as a QUBO minimization problem, with the long-term goal of realizing a fully quantum annealing-based spectral solver with minimal hybrid overhead.

Acknowledgments

G.D.G. acknowledges the support from the EU-FESR, PON Ricerca e Innovazione, Grant No. 2014-2020-DM 1062/2021. This work is also supported the Czech Science Foundation (Czech Republic), P203-26-21972S. Computational resources were partially provided by the e-INFRA CZ project (ID:90254), supported by the Ministry of Education, Youth and Sports of the Czech Republic, and by the ELIXIR-CZ project (ID:90255), supported by the Ministry of Education, Youth and Sports of the Czech Republic, and by the ELIXIR-CZ project (ID:90255).

References

  1. [31]A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’Brien, A variational eigenvalue solver on a photonic quantum processor, Nature Communications 5, 4213 (2014).DOI
  2. [32]J. Tilly, H. Chen, S. Cao, D. Picozzi, K. Setia, Y. Li, E. Grant, L. Wossnig, I. Rungger, G. H. Booth, and J. Tennyson, The variational quantum eigensolver: A review of methods and best practices, Physics Reports 986, 1 (2022), the Variational Quantum Eigensolver: a review of methods and best practices.DOI
  3. [33]W. Kirby, M. Motta, and A. Mezzacapo, Exact and efficient Lanczos method on a quantum computer, Quantum 7, 1018 (2023).DOI
  4. [34]K. M. Nakanishi, K. Mitarai, and K. Fujii, Subspace-search variational quantum eigensolver for excited states, Phys. Rev. Res. 1, 033062 (2019).DOI
  5. [35]O. Higgott, D. Wang, and S. Brierley, Variational quantum eigensolver for excited states, Quantum 3, 156 (2019).
  6. [36]R. M. Parrish, E. G. Hohenstein, P. L. McMahon, and T. J. Martínez, Quantum computation of electronic transitions using a variational quantum eigensolver, Phys. Rev. Lett. 122, 230401 (2019).DOI
  7. [37]N. V. Tkachenko, L. Cincio, A. I. Boldyrev, S. Tretiak, P. A. Dub, and Y. Zhang, Quantum davidson algorithm for excited states, Quantum Science and Technology 9, 035012 (2024).arxiv.org/abs/2204.10741
  8. [38]J. R. McClean, M. E. Kimchi-Schwartz, J. Carter, and W. A. de Jong, Hybrid quantum-classical hierarchy for mitigation of decoherence and determination of excited states, Phys. Rev. A 95, 042308 (2017).
  9. [39]J. I. Colless, V. V. Ramasesh, D. Dhlen, M. S. Blok, M. E. Kimchi-Schwartz, J. R. McClean, J. Carter, W. A. de Jong, and I. Siddiqi, Computation of molecular spectra on a quantum processor with an error-resilient algorithm, Phys. Rev. X 8, 011021 (2018).DOI
  10. [40]T. Takeshita, N. C. Rubin, Z. Jiang, E. Lee, R. Babbush, and J. R. McClean, Increasing the representation accuracy of quantum simulations of chemistry without extra quantum resources, Phys. Rev. X 10, 011004 (2020).
  11. [41]W. J. Huggins, J. Lee, U. Baek, B. O’Gorman, and K. B. Whaley, A non-orthogonal variational quantum eigensolver, New Journal of Physics 22, 073009 (2020).DOI
  12. [42]P. J. Ollitrault, A. Kandala, C.-F. Chen, P. K. Barkoutsos, A. Mezzacapo, M. Pistoia, S. Sheldon, S. Woerner, J. M. Gambetta, and I. Tavernelli, Quantum equation of motion for computing molecular excitation energies on a quantum computer, Phys. Rev. Res. 2, 043140 (2020).
  13. [43]A. Asthana et al., Equation-of-motion based methods for quantum computing: excited states and beyond, Chemical Science 14, 2405 (2023).
  14. [44]A. Kumar, A. Asthana, V. Abraham, T. D. Crawford, N. J. Mayhall, Y. Zhang, L. Cincio, S. Tretiak, and P. A. Dub, Variational quantum eigensolver for excited states using equation-of-motion ansatz, Journal of Chemical Theory and Computation 19, 9136 (2023).
  15. [45]M. Q. Hlatshwayo, J. Novak, and E. Litvinova, Quantum benefit of the quantum equation of motion for the strongly coupled many-body problem, Phys. Rev. C 109, 014306 (2024).
  16. [46]C. L. Cortes and S. K. Gray, Quantum krylov subspace algorithms for quantum dynamics simulations, Phys. Rev. A 105, 022417 (2022).
  17. [47]N. H. Stair, R. Huang, and F. A. Evangelista, Simulating many-body quantum chemistry with quantum computers, Journal of Chemical Theory and Computation 16, 2236 (2020).
  18. [48]M. Motta, C. Sun, A. T. K. Tan, M. J. O’Rourke, E. Ye, A. J. Minnich, F. G. S. L. Brandão, and G. K.-L. Chan, Determining eigenstates and thermal states on a quantum computer using quantum imaginary time evolution, Nat. Phys. 16, 205 (2020).
  19. [49]B. Krakoff, S. M. Mniszewski, and C. F. A. Negre, Controlled precision qubo-based algorithm to compute eigenvectors of symmetric matrices, PLOS ONE 17, 1 (2022).
  20. [50]M. Illa and M. J. Savage, Basic elements for simulations of standard-model physics with quantum annealers: Multigrid and clock states, Physical Review A 106, 052605 (2022).
  21. [51]M. G. Balducci, B. Chen, M. Möller, M. Gerritsma, and R. D. Breuker, Review and perspectives in quantum computing for partial differential equations in structural mechanics, Frontiers in Mechanical Engineering 8, 914241 (2022).
  22. [52]A. Stockley and K. Briggs, Optimizing antenna beam-forming with quantum computing, in Proceedings of the 2023 17th European Conference on Antennas and Propagation (EuCAP) (IEEE, 2023) pp. 1–5.
  23. [53]A. Rajak, S. Suzuki, A. Dutta, and B. K. Chakrabarti, Quantum annealing: an overview, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 381, 20210417 (2023).DOI
  24. [54]F. Fornetti, A. Gnec, T. Frederico, F. Pedretti, M. Rinaldi, A. Roggero, G. Salmè, S. Scopetta, and M. Viviani, Solving the homogeneous bethe-salpeter equation with a quantum annealer, Phys. Rev. D 110, 056012 (2024).arxiv.org/abs/2406.18669
  25. [55]H. Wolkowicz and G. P. H. Styan, Bounds for eigenvalues using traces, Linear Algebra and its Applications 29, 471 (1980).DOI
  26. [56]J. Nocedal and S. J. Wright, Numerical Optimization, 2nd ed. (Springer, New York, 2006).
  27. [57]C. B. Moler and G. W. Stewart, An algorithm for generalized matrix eigenvalue problems, SIAM Journal on Numerical Analysis 10, 241 (1973).DOI
  28. [58]B. Kågström and D. Kressner, Multishift variants of the qz algorithm with aggressive early deflation, SIAM Journal on Matrix Analysis and Applications 29, 199 (2006).
  29. [59]J. G. F. Francis, The qr transformation. a unitary analogue to the lr transformation—part 1, The Computer Journal 4, 265 (1961).
  30. [60]J. G. F. Francis, The qr transformation—part 2, The Computer Journal 4, 332 (1962).
  31. [61]A. Shirokov, I. Shin, Y. Kim, M. Sosonkina, P. Maris, and J. Vary, N3LO NN interaction adjusted to light nuclei in ab exit[u] approach, Physics Letters B 761, 87 (2016).
  32. [62]F. Wegner, Flow-equations for hamiltonians, Annalen der Physik 506, 77 (1994).
  33. [63]D.-W. Quantum, The advantage2 quantum computer systems, https://www.dwavequantum.com/solutions-and-products/systems/ (2025).
  34. [64]G. W. Stewart, Matrix Algorithms, Volume II: Eigensystems (SIAM, 2001).
  35. [65]C. F. Mehl and P. V. Dooren, On the stability of deflation techniques for generalized eigenvalue problems, SIAM Journal on Matrix Analysis and Applications 24, 658 (2003).
  36. [66]D. S. Watkins, The Matrix Eigenvalue Problem: GR and Krylov Subspace Methods (SIAM, 2007).

Paper details

vlnc:W_9v3wye5y
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents