betaFeedback

WorksPhysics and Astronomy

Constrained Quantum Imaginary Time in Krylov Space

Abstract

We introduce constrained quantum imaginary time evolution (CQITE), extending quantum imaginary time evolution (QITE) to constrained-search dynamics for ground-state preparation with prescribed expectation-value constraints. A central challenge in practical QITE is constructing a compact unitary representation of the imaginary-time evolution while maintaining accuracy and controlling the approximation error. We resolve this challenge by deriving an exact characterization of the imaginary-time evolution from the Krylov space generated by repeated applications of the Hamiltonian to the initial state. This characterization induces a hierarchy of operator directions expressed through nested anti-commutators. Truncating the hierarchy produces a systematic static operator pool, with polynomial size in the system size at any fixed truncation order. The resulting construction provides a principled route to compact unitary approximations of imaginary-time dynamics. Numerical tests on spin models show close agreement with exact constrained imaginary-time evolution and small constraint deviations, while a fermionic example yields energies close to the analytic constrained-search reference.

Introduction

Preparing ground states of quantum many-body systems is a central task in quantum simulation, quantum chemistry, and combinatorial optimization. Imaginary time evolution provides an effective route toward ground-state preparation by exponentially suppressing excited-state components. However, its application to large-scale quantum systems is hindered by two fundamental obstacles: the exponential growth of the many-body Hilbert space makes exact classical simulations intractable, while its intrinsically non-unitary dynamics prevents direct implementation on gate-based quantum computers [5, 6, 25, 24, 26].

Quantum imaginary time evolution (QITE) [3] addresses this limitation by replacing infinitesimal imaginary-time steps with locally optimal unitary transformations that reproduce the imaginary-time dynamics. This hybrid quantum–classical framework has emerged as a promising approach for ground-state preparation on gate-based quantum devices through shallow unitary updates [20, 6, 7].

In many applications, however, the target ground state is required to satisfy additional physical constraints, such as prescribed expectation values of conserved observables, symmetry conditions, or feasibility requirements arising in constrained optimization problems [10, 11, 2, 1]. Constrained-search imaginary time evolution extends the imaginary-time formalism by introducing time-dependent Lagrange multipliers that enforce such constraints throughout the evolution [1]. Since the resulting dynamics remain intrinsically non-unitary, their direct realization on gate-based quantum hardware remains an open challenge.

In this work we introduce the constrained quantum imaginary time evolution (CQITE) algorithm, extending the standard QITE algorithm to constrained-search dynamics.

The implementation of both QITE and CQITE requires the definition of a unitary operator that mimics the imaginary-time dynamics. In practice, this requires a finite support. Although the full Pauli support can represent the exact dynamics, its size grows exponentially with the system size. It is therefore not evident how to construct a compact operator support without sacrificing accuracy. Existing approaches can construct the support adaptively, but they do not directly characterize the operator directions generated by the imaginary-time evolution [8, 9].

We address this problem by deriving an exact characterization of the imaginary-time evolution based on a Krylov space generated from the initial state through repeated anti-commutator actions of the Hamiltonian.

Unlike conventional quantum Krylov methods, which construct Krylov spaces from Hamiltonian powers or real-time evolution [15, 16, 17, 18], the Krylov hierarchy derived here follows directly from the operator structure of imaginary-time evolution. Truncating this hierarchy yields a static operator whose size scales polynomially with the system size for fixed truncation order. Increasing the truncation order enlarges the accessible space and systematically approaches the full Krylov space. The resulting construction provides an efficient alternative to variational operator pools that can be applied to both unconstrained and constrained quantum imaginary-time evolution. The resulting framework therefore provides two complementary advances. First, it extends quantum imaginary-time evolution to constrained-search dynamics, yielding a practical algorithm for constrained ground-state preparation. Second, it establishes an exact Krylov characterization of the operator directions generated by imaginary-time evolution. The remainder of the paper is organized as follows. § 2 reviews constrained imaginary-time evolution and the constrained-search formalism underlying the proposed approach. § 3 derives the CQITE algorithm. § 4 develops the Krylov hierarchy. § 5 presents numerical validation. Technical details concerning the preparation of the constrained reference state, the polynomial approximation of the density operator and the preservation of the Krylov support under constraints are collected in the appendices.

Algorithm 1 Constrained Quantum Imaginary Time Evolution

input: Hamiltonian HH, constraint C\mathbf{C}, target b\mathbf{b} ▷ Pauli data are efficiently encoded

  1. hierarchy P←vec⁡⋃jsupp⁡Hj\mathbf{P} \leftarrow\operatorname{vec}\bigcup_j \operatorname{supp} H^j ▷ § 4

  2. reference ∣φ(0)⟩←|\varphi(0)\rangle\leftarrow any ∣φ⟩|\varphi\rangle s.t. ⟨C⟩=b\langle\mathbf{C}\rangle=\mathbf{b}, τ←0\tau\leftarrow0 ▷ Sec. A

  3. repeat

  4. evolve ▷ Run multiple times for sampling

∣φ(τ)⟩←Texp⁡[−i∫0τdς A(ς)]∣φ(0)⟩|\varphi(\tau)\rangle\leftarrow\mathcal{T}\exp\left[-\mathrm{i}\int_{0}^{\tau}\mathrm{d}\varsigma\,A(\varsigma)\right]|\varphi(0)\rangle
  1. estimate Eq. (5) and Eq. (14)

  2. enforce v←−⟨⟨CCT⟩⟩−1⟨⟨HC⟩⟩\mathbf{v}\leftarrow-\langle\langle\mathbf{C}\mathbf{C}^{\mathrm{T}}\rangle\rangle^{-1}\langle\langle H\mathbf{C}\rangle\rangle ▷ § 2

  3. project a←−Re⁡⟨PPT⟩−1Im⁡⟨P(H+CTv)⟩\mathbf{a}\leftarrow-\operatorname{Re}\langle\mathbf{P}\mathbf{P}^{\mathrm{T}}\rangle^{-1}\operatorname{Im}\langle\mathbf{P}(H+\mathbf{C}^{\mathrm{T}}\mathbf{v})\rangle ▷ § 3

  4. update A∣(τ,τ+δ]←aTPA\big|_{(\tau,\tau+\delta]}\leftarrow\mathbf{a}^{\mathrm{T}}\mathbf{P}, τ←τ+δ\tau\leftarrow\tau+\delta

  5. until Eq. (6)

  6. return ⟨H⟩\langle H\rangle

Imaginary time under constraints

Let

C=(C1,…,Cm)T∈span⁡Pwm.(1)\mathbf{C}=(C_1,\ldots,C_m)^{\mathrm{T}}\in\operatorname{span}\mathcal{P}_w^m . \tag*{(1)}

be a vector of constraints, where each CjC_j is expressed as a linear combination of operators coming from the Pauli group Pw={P1⊗⋯⊗Pw∣P∈{1,X,Y,Z}}\mathcal{P}_w=\{P_1\otimes\cdots\otimes P_w\mid P\in\{1,X,Y,Z\}\}. Given a target vector b∈Rm\mathbf{b}\in\mathbb{R}^m, we consider the constrained optimization problem

min⁡∣φ⟩⟨H⟩s.t.⟨C⟩=b,(2)\begin{aligned} \min_{|\varphi\rangle}\quad&\langle H\rangle\\ \text{s.t.}\quad&\langle\mathbf{C}\rangle=\mathbf{b}, \tag*{(2)} \end{aligned}

with HH being a linear combination of nn operators in Pw\mathcal{P}_w. Following the constrained-search formalism [10, 11] recently extended to imaginary time [1], the optimization method is carried out by solving the differential equation

−∂τ∣φ(τ)⟩=(G(τ)−⟨G(τ)⟩)∣φ(τ)⟩,G(τ)=H+v(τ)TC,(3)-\partial_{\tau}|\varphi(\tau)\rangle= \left(G(\tau)-\langle G(\tau)\rangle\right)|\varphi(\tau)\rangle,\qquad G(\tau)=H+\mathbf{v}(\tau)^{T}\mathbf{C}, \tag*{(3)}

where v∈Rm\mathbf{v}\in\mathbb{R}^{m} is the vector of real time-dependent Lagrange multipliers enforcing the constraints. In the limit τ→∞\tau\to\infty the term vTC\mathbf{v}^{T}\mathbf{C} represents an external potential which added to the Hamiltonian ensures that the ground state of the system in the presence of this extra potential v\mathbf{v} has the prescribed expectation values. It can thus be seen as control potential that can adjust ground states so that they have desired expectation values. A good physical example is to tune an external magnetic field for a spin-lattice until a desired spin-magnetization expectation value is achieved in the ground state.

Preservation of the constraints requires

ddτ⟨C⟩=0.(4)\frac{\mathrm{d}}{\mathrm{d}\tau}\langle\mathbf{C}\rangle=0. \tag*{(4)}

Using Eq. (3), the Lagrange multipliers are determined by the real linear system

⟨ ⁣⟨CCT⟩ ⁣⟩v=−⟨ ⁣⟨HC⟩ ⁣⟩,(5)\langle\!\langle\mathbf{C}\mathbf{C}^{T}\rangle\!\rangle\mathbf{v}=-\langle\!\langle H\mathbf{C}\rangle\!\rangle, \tag*{(5)}

where ⟨ ⁣⟨⋅⟩ ⁣⟩\langle\!\langle\cdot\rangle\!\rangle denotes the symmetrized covariance. The matrix ⟨ ⁣⟨CCT⟩ ⁣⟩\langle\!\langle\mathbf{C}\mathbf{C}^{T}\rangle\!\rangle is state-dependent and it is not guaranteed to be invertible for every state ∣φ⟩|\varphi\rangle. If the constraint matrix is singular, the multipliers are understood in the Moore–Penrose pseudoinverse sense.

At convergence, the flow reaches a stationary point satisfying

(G−⟨G⟩)∣φ⟩=0.(6)\left(G-\langle G\rangle\right)|\varphi\rangle=0. \tag*{(6)}

This condition characterizes a critical point of the energy functional, and it can be shown that such a stationary point corresponds to an absolute minimizer under the constraints [1], provided that ∣φ(0)⟩|\varphi(0)\rangle has a non-zero overlap with it.

Constrained QITE

We solve the imaginary time equation Eq. (3) by assuming that the generator is piecewise constant in δ\delta. The step generator GG is built from the incoming state, according to Eq. (5). For an incoming state ∣φ(τ−δ)⟩|\varphi(\tau-\delta)\rangle, the exact normalized constrained imaginary time step is given by

∣φ⋆(τ)⟩=e−δG⟨e−2δG⟩∣φ(τ−δ)⟩.(7)|\varphi^\star(\tau)\rangle=\frac{e^{-\delta G}}{\sqrt{\langle e^{-2\delta G}\rangle}}|\varphi(\tau-\delta)\rangle. \tag*{(7)}

The exact step Eq. (7) is non-unitary. Since both ∣φ(τ−δ)⟩|\varphi(\tau-\delta)\rangle and ∣φ⋆(τ)⟩|\varphi_\star(\tau)\rangle are normalized, there exists at least one unitary operator U⋆U_\star such that

∣φ⋆(τ)⟩=U⋆∣φ(τ−δ)⟩(8)|\varphi_\star(\tau)\rangle= U_\star|\varphi(\tau-\delta)\rangle \tag*{(8)}

under the piecewise constant assumption. The CQITE step seeks a structured Hermitian operator AA such that

∣φ(τ)⟩=e−iδA∣φ(τ−δ)⟩(9)|\varphi(\tau)\rangle= e^{-i\delta A}|\varphi(\tau-\delta)\rangle \tag*{(9)}

approximates the exact step Eq. (7) locally in δ\delta.

Operator expansion and variational equations

To construct a practical approximation to the normalized imaginary time evolution, we expand the unitary generator in a restricted operator pool P=(P1,…,Pw)T\mathbf{P} = (P_1,\ldots,P_w)^{\mathsf{T}}, and coefficients a=(a1,…,aw)T\mathbf{a} = (a_1,\ldots,a_w)^{\mathsf{T}}, for some w≥1w \ge1.

The effective generator is written as

A=aTP,(10)A = \mathbf{a}^{\mathsf{T}}\mathbf{P}, \tag*{(10)}

where the coefficients are determined variationally [20, 6, 7]. Expanding the normalized imaginary time step to first order in δ\delta gives

∣φ⋆(τ)⟩=∣φ(τ−δ)⟩−δ(G−⟨G⟩)∣φ(τ−δ)⟩+O(δ2),(11)|\varphi_\star(\tau)\rangle= |\varphi(\tau-\delta)\rangle- \delta\left(G-\langle G\rangle\right)|\varphi(\tau-\delta)\rangle+ O(\delta^2), \tag*{(11)}

Expanding the variational unitary ansatz to first order in δ\delta yields

∣φ(τ)⟩=∣φ(τ−δ)⟩−iδA∣φ(τ−δ)⟩+O(δ2).(12)|\varphi(\tau)\rangle= |\varphi(\tau-\delta)\rangle- i\delta A|\varphi(\tau-\delta)\rangle+ O(\delta^2). \tag*{(12)}

The coefficients of AA are obtained by requiring stationarity of the squared distance between the exact tangent vector and its variational approximation [23, 6, 5],

∇a⟨φ⋆(τ)−φ(τ)∣φ⋆(τ)−φ(τ)⟩=0.(13)\nabla_{\mathbf{a}}\langle\varphi_\star(\tau)-\varphi(\tau)|\varphi_\star(\tau)-\varphi(\tau)\rangle= 0. \tag*{(13)}

Matching the first-order terms in the two expansions projects the exact imaginary time update onto the variational subspace spanned by ∇a∣φ(τ−δ)⟩\nabla_{\mathbf{a}}|\varphi(\tau-\delta)\rangle and yields the linear system

Re⁡⟨PPT⟩a=−Im⁡⟨PG⟩.(14)\operatorname{Re}\langle\mathbf{P}\mathbf{P}^{\mathsf{T}}\rangle\mathbf{a} = -\operatorname{Im}\langle\mathbf{P}G\rangle. \tag*{(14)}

State-dependent occupation number preservation

A large class of Hamiltonians of physical interest is that of interacting fermions, given in second quantization notation by the general form

H=∑p,qhpqcp†cq+∑p,q,s,rvpqsrcp†cq†cscr(15)H = \sum_{p,q} h_{pq}c_p^\dagger c_q + \sum_{p,q,s,r} v_{pqsr}c_p^\dagger c_q^\dagger c_s c_r \tag*{(15)}

where cp†c_p^\dagger and cpc_p are fermionic creation an annihilation operators. This class of Hamiltonians includes the Hubbard model or the general form of a molecular Hamiltonian in quantum chemistry. An important class of constraints is that represented by one-body operators of the form

Cj=∑p,qDpq(j)cp†cq(16)C_j = \sum_{p,q} D_{pq}^{(j)} c_p^\dagger c_q \tag*{(16)}

for Hermitian matrices Dpq(j)D_{pq}^{(j)}. As an example of a set of commuting constraints we can take the site occupations nj=cj†cjn_j = c_j^\dagger c_j of the Hubbard model or of lattice density-functional theory. Using a qubit encoding of the fermionic Fock space, such as the Jordan–Wigner transformation, the vector of site-occupation operators is represented as

n=1−Z2(17)\mathbf{n} = \frac{1-\mathbf{Z}}{2} \tag*{(17)}

where

n=(c1†c1,…,cm†cm)T,Z=(Z1,…,Zm)T=C(18)\mathbf{n} = \left(c_1^\dagger c_1,\ldots,c_m^\dagger c_m\right)^{\mathrm{T}}, \qquad\mathbf{Z} = \left(Z_1,\ldots,Z_m\right)^{\mathrm{T}} = \mathbf{C} \tag*{(18)}

are the vectors of occupation-number operators and constraints, respectively, for mm fermionic modes. The corresponding occupation expectation values are therefore given by

⟨n⟩=1−⟨C⟩2.(19)\langle\mathbf{n}\rangle= \frac{1-\langle\mathbf{C}\rangle}{2}. \tag*{(19)}

Thus, the admissible target values satisfy

−1≤b≤1,(20)-1 \le\mathbf{b} \le1, \tag*{(20)}

which is equivalent to requiring

0≤⟨n⟩≤1.(21)0 \le\langle\mathbf{n}\rangle\le1. \tag*{(21)}

The target values are chosen consistently with the desired nn-particle sector, i.e.

1Tb=m−2n.(22)\mathbf{1}^{\mathrm{T}}\mathbf{b} = m - 2n. \tag*{(22)}

The initial state must be prepared using number-conserving two-mode rotations. Since both the Hamiltonian and the constraints preserve particle number, the constrained imaginary-time evolution of Eq. (7) remains in the same particle-number sector. The CQITE unitary is then constructed state-dependently to reproduce this constrained imaginary-time update. Crucially, particle-number preservation is therefore required only for the action of the unitary on the state generated by the algorithm, and not as an operator-level symmetry of the unitary itself. The unitary need not preserve particle number for arbitrary input states: it is sufficient that, for the particular input state encountered during the evolution, it reproduces the corresponding constrained imaginary-time state, which belongs to the same particle-number sector. Particle-number preservation thus follows from the appropriately initialized state and the constrained imaginary-time evolution. However, the approximating nature of the finite operator pool may result in particle-number leakage in the practical implementation.

Static polynomial Krylov spaces

The CQITE formulation derived in the previous section is exact provided that the unitary generator is expanded in the complete Pauli operator basis. In practice, however, such an expansion is computationally prohibitive, since the number of Pauli operators grows exponentially with the system size. The central challenge is therefore to identify a compact operator pool that is sufficiently expressive to reproduce the (constrained) imaginary-time dynamics while scaling efficiently with system size.

The structure of the exact QITE generator provides a systematic route to this construction. Rather than selecting the operator pool heuristically, we derive it directly from the operator content of the exact imaginary-time evolution. We show that the relevant operator directions form a hierarchy of Krylov spaces generated by repeated actions of the Hamiltonian on the operator space. Truncating this hierarchy at a fixed order yields a static operator pool whose size grows polynomially with the system size.

The Krylov sequence provides a systematic characterization of the energy-driven directions accessible from the reference state. We therefore derive the corresponding operator hierarchy for the Hamiltonian alone, for simplicity. The resulting operator pools can then be employed in the constrained setting, where they perform well numerically, as demonstrated in § 5.

Our proposal is closely related to the broader idea of quantum Krylov subspace methods, which have recently emerged as an alternative paradigm for quantum eigensolvers based on Hamiltonian powers or real-time evolution [16, 15, 17, 18, 19]. Here, however, the Krylov hierarchy is not introduced as an independent variational ansatz, but follows directly from the nested anti-commutator structure of the exact imaginary-time propagator. Truncating this hierarchy therefore provides a systematic sequence of approximations with increasing expressive power.

A unitary representation for imaginary time

Let us define the Hermitian operator

A(τ)=−i[G,Φ(τ)]=A†(τ),(23)A(\tau) = -i [G, \Phi(\tau)] = A^\dagger(\tau), \tag*{(23)}

with Φ(τ)=∣φ(τ)⟩⟨φ(τ)∣\Phi(\tau) = |\varphi(\tau)\rangle\langle\varphi(\tau)|.

Applying AA to the state gives

A(τ)∣φ(τ)⟩=−i(GΦ(τ)−Φ(τ)G)∣φ(τ)⟩=−i(G−⟨G⟩)∣φ(τ)⟩.(24)A(\tau)|\varphi(\tau)\rangle= -i\left(G\Phi(\tau) - \Phi(\tau)G\right)|\varphi(\tau)\rangle= -i\left(G-\langle G\rangle\right)|\varphi(\tau)\rangle. \tag*{(24)}

Therefore,

i∂τ∣φ(τ)⟩=A(τ)∣φ(τ)⟩,(25)i\partial_\tau|\varphi(\tau)\rangle= A(\tau)|\varphi(\tau)\rangle, \tag*{(25)}

which has the form of a Schrödinger equation. A key observation is that A(τ)A(\tau) depends on the very state on which it acts, but we will suppress this dependence in the notation for clarity of presentation. Moreover, even if GG is a one-body or two-body operator, AA is generally an nn-body operator.

The Schrödinger equation can be rewritten in integral form as

∣φ(τ)⟩=∣φ(0)⟩−i∫0τdς A(ς)∣φ(ς)⟩.(26)|\varphi(\tau)\rangle= |\varphi(0)\rangle- i\int_{0}^{\tau} \mathrm{d}\varsigma\, A(\varsigma)|\varphi(\varsigma)\rangle. \tag*{(26)}

This naturally suggests the time-stepping scheme

∣φ(τ+δ)⟩=∣φ(τ)⟩−i∫ττ+δdς A(ς)∣φ(ς)⟩≈e−iδA(τ)∣φ(τ)⟩(27)|\varphi(\tau+\delta)\rangle= |\varphi(\tau)\rangle- i\int_{\tau}^{\tau+\delta} \mathrm{d}\varsigma\, A(\varsigma)|\varphi(\varsigma)\rangle\approx e^{-i\delta A(\tau)}|\varphi(\tau)\rangle \tag*{(27)}

for sufficiently small δ\delta. For τ→∞\tau\to\infty, this evolution converges to the ground state provided the initial state has a nonzero overlap with the ground state.

Krylov representation of the evolution

For simplicity, we restrict the analysis to the Hamiltonian HH, for which we derive the exact propagator. The QITE propagator satisfies

∣φ(τ+δ)⟩≈e−iδA(τ)∣φ(τ)⟩=e−δ[H,Φ(τ)]∣φ(τ)⟩.(28)|\varphi(\tau+\delta)\rangle\approx e^{-i\delta A(\tau)}|\varphi(\tau)\rangle= e^{-\delta[H,\Phi(\tau)]}|\varphi(\tau)\rangle. \tag*{(28)}

To characterize the operator content of the propagator, we express the density operator directly from the normalized imaginary time evolution. The solution of

−∂τ∣φ(τ)⟩=(H−⟨H⟩)∣φ(τ)⟩(29)-\partial_{\tau}|\varphi(\tau)\rangle= (H-\langle H\rangle)|\varphi(\tau)\rangle \tag*{(29)}

is

∣φ(τ)⟩=e−τH∣φ(0)⟩⟨φ(0)∣e−2τH∣φ(0)⟩.(30)|\varphi(\tau)\rangle= \frac{e^{-\tau H}|\varphi(0)\rangle}{\sqrt{\langle\varphi(0)|e^{-2\tau H}|\varphi(0)\rangle}} . \tag*{(30)}

The associated density operator takes the form

Φ(τ)=e−τHΦ(0)e−τHTr⁡(e−2τHΦ(0)).(31)\Phi(\tau) = \frac{e^{-\tau H}\Phi(0)e^{-\tau H}}{\operatorname{Tr}(e^{-2\tau H}\Phi(0))}. \tag*{(31)}

Since the denominator is a scalar, it factors out of the commutator:

[H,Φ(τ)]=e−τH[H,Φ(0)]e−τHTr⁡(e−2τHΦ(0))=ie−τHA(0)e−τHTr⁡(e−2τHΦ(0)).(32)[H,\Phi(\tau)] = \frac{e^{-\tau H}[H,\Phi(0)]e^{-\tau H}}{\operatorname{Tr}(e^{-2\tau H}\Phi(0))} = i\frac{e^{-\tau H}A(0)e^{-\tau H}}{\operatorname{Tr}(e^{-2\tau H}\Phi(0))}. \tag*{(32)}

The normalization factor only rescales the generator and does not modify its operator support. Consequently, the operator directions are entirely determined by

A(τ)∝e−τHA(0)e−τH.(33)A(\tau) \propto e^{-\tau H}A(0)e^{-\tau H}. \tag*{(33)}

The imaginary-time evolution satisfies

e−τHA(0)e−τH=e−τac⁡HA(0),(34)e^{-\tau H}A(0)e^{-\tau H}=e^{-\tau\operatorname{ac}_{H}}A(0), \tag*{(34)}

where ac⁡H\operatorname{ac}_{H} denotes the anti-commutator operator ac⁡HA=HA+AH\operatorname{ac}_{H} A=HA+AH.

Expanding the exponential,

e−τHA(0)e−τH=∑j∞(−τ)jj!ac⁡HjA(0).(35)e^{-\tau H}A(0)e^{-\tau H}=\sum_{j}^{\infty}\frac{(-\tau)^{j}}{j!}\operatorname{ac}_{H}^{j}A(0). \tag*{(35)}

Corollary. The operator directions are spanned by the Krylov subspace [12, 13]

K(ac⁡H,A(0))=span⁡{ac⁡HjA(0)}j,dim⁡K<∞(36)\mathcal{K}(\operatorname{ac}_{H},A(0))=\operatorname{span}\left\{\operatorname{ac}_{H}^{j}A(0)\right\}_{j},\quad\dim\mathcal{K}<\infty \tag*{(36)}

The operator pool is therefore constructed from the Pauli support of the Krylov sequence. This choice is directly motivated by the structure of the exact imaginary time evolution and provides a systematic hierarchy of approximations. Increasing the truncation order enlarges the accessible variational manifold and converges toward the full Krylov space generated by the anticommutator superoperator.

The resulting pool provides a set of operators that captures the dominant directions explored by the exact dynamics while remaining independent of the instantaneous values of the variational coefficients.

Polynomial construction of the pool

The construction above defines the Krylov pool through the operator sequence {ac⁡HjA(0)}j\{\operatorname{ac}_{H}^{j}A(0)\}_{j}, but a direct Pauli expansion of A(0)A(0) would still require an explicit Pauli representation of Φ(0)\Phi(0). The dyadic structure of the initial state allows us to avoid this expansion altogether and construct a polynomial-size superset of the same Krylov directions directly from powers of HH.

The operator Krylov sequence action satisfies

ac⁡HjA(0)∣φ(0)⟩∈span⁡{∣φ(0)⟩,H∣φ(0)⟩,…,Hj+1∣φ(0)⟩},(37)\operatorname{ac}_{H}^{j}A(0)\lvert\varphi(0)\rangle\in\operatorname{span}\left\{\lvert\varphi(0)\rangle,H\lvert\varphi(0)\rangle,\ldots,H^{j+1}\lvert\varphi(0)\rangle\right\}, \tag*{(37)}

This relation allows the Pauli pool to be constructed without expanding Φ(0)\Phi(0). For a Hamiltonian composed by nn Pauli operators, each product is itself Pauli up to a phase. The number of distinct Pauli strings in the support of HjH^{j} is bounded by

∣supp⁡Hj∣≤nj.(38)\left|\operatorname{supp} H^{j}\right|\le n^{j}. \tag*{(38)}

We therefore define the Pauli-support vector

P=vec⁡⋃jcsupp⁡Hj,(39)\mathbf{P}=\operatorname{vec}\bigcup_{j}^{c}\operatorname{supp} H^{j}, \tag*{(39)}

with vec⁡{P1,…,Pw}=(P1,…,Pw)T\operatorname{vec}\{P_{1},\ldots,P_{w}\}=(P_{1},\ldots,P_{w})^{\mathsf{T}}.

The resulting pool contains all Pauli directions required by the truncated Krylov construction. Importantly, the imaginary-time parameter τ\tau does not enter the construction of the Pauli pool. Hence, τ\tau determines only the coefficients multiplying the Krylov directions, while the Pauli pool itself is determined solely by the Krylov order cc and the Hamiltonian. Therefore, the pool can be constructed once and subsequently used for all values of τ\tau. The size of this pool is bounded by O(nc)O(n^c). No Pauli expansion of Φ(0)\Phi(0) is required.

Norm and truncation bounds

Using the Krylov expansion of the exact generator, the contribution associated with the first cc Krylov operators is contained in P\mathbf{P} of Eq. (39), which is a vector of Pauli operators

P=(P1,…,PO(nc))T.(40)\mathbf{P} = (P_1,\ldots,P_{O(n^c)})^{\mathsf{T}}. \tag*{(40)}

Assuming that the variational coefficients a\mathbf{a} are obtained exactly, the remaining error is therefore bounded by the contribution of the Krylov tail. Hence,

∥(A(τ)−aTP)∣φ(τ)⟩∥2≲∥∑j>c(−τ)jj!HjA(0)∣φ(τ)⟩∥2.(41)\left\lVert\left(A(\tau)-\mathbf{a}^{\mathsf{T}}\mathbf{P}\right)|\varphi(\tau)\rangle\right\rVert_2 \lesssim\left\lVert\sum_{j>c}\frac{(-\tau)^j}{j!}\mathcal{H}^{j}A(0)|\varphi(\tau)\rangle\right\rVert_2. \tag*{(41)}

The state dependence can be removed from the upper bound by using

∥HHcA(0)∣φ(τ)⟩∥2≤∥HHcA(0)∥≤2∥Hj+1∥.(42)\left\lVert\mathcal{H}^{c}_{H}A(0)|\varphi(\tau)\rangle\right\rVert_2 \leq\left\lVert\mathcal{H}^{c}_{H}A(0)\right\rVert\leq2\lVert H^{j+1}\rVert. \tag*{(42)}

Thus,

∥(A(τ)−aTP)∣φ(τ)⟩∥2≲∑j>c(τ∥H∥)jj!.(43)\left\lVert\left(A(\tau)-\mathbf{a}^{\mathsf{T}}\mathbf{P}\right)|\varphi(\tau)\rangle\right\rVert_2 \lesssim\sum_{j>c}\frac{(\tau\lVert H\rVert)^j}{j!}. \tag*{(43)}

Finally, applying the standard upper bound for the remainder of the exponential series, gives the uniform bound

∥(A(τ)−aTP)∣φ(τ)⟩∥2≲(τ∥H∥)c+1(c+1)!(44)\left\lVert\left(A(\tau)-\mathbf{a}^{\mathsf{T}}\mathbf{P}\right)|\varphi(\tau)\rangle\right\rVert_2 \lesssim\frac{(\tau\lVert H\rVert)^{c+1}}{(c+1)!} \tag*{(44)}

Since the residual bound is monotonically increasing with τ\tau, it is sufficient to evaluate the bound at the characteristic convergence timescale set by the spectral gap, i.e. τ≲1E1−E0\tau\lesssim\frac{1}{E_1-E_0}, and normalized by ∥H∥\lVert H\rVert. Hence we obtain

∥(A(τ)−aTP)∣φ(τ)⟩∥2≲1(c+1)!Δc+1,(45)\left\lVert\left(A(\tau)-\mathbf{a}^{\mathsf{T}}\mathbf{P}\right)|\varphi(\tau)\rangle\right\rVert_2 \lesssim\frac{1}{(c+1)!\Delta^{c+1}}, \tag*{(45)}

with Δ=(E1−E0)/∥H∥\Delta=(E_1-E_0)/\lVert H\rVert, satisfying 0<Δ≤20<\Delta\leq2. Hence, within the fast-convergence regime Δ≥1\Delta\geq1 the residual exhibits a factorial decay with cc. For Δ<1\Delta<1, a larger Krylov order may be required to achieve the same residual accuracy, e.g.

c=Θ(Δ−1),(46)c=\Theta(\Delta^{-1}), \tag*{(46)}

namely, cc is asymptotically bounded both below and above by Δ−1\Delta^{-1}. A favorable scenario is given by Δ=Θ(1)\Delta= \Theta(1), where cc can be kept constant and the pool size therefore scales polynomially with nn.

For gaps that close sufficiently rapidly, this state-independent bound may become increasingly conservative, as it does not account for additional structure that can reduce the required Krylov order.

The above derivation also allows us to obtain a dimensionless truncation measure bounded by unity. We normalize the Taylor tail by the total weight of the exponential series. In fact, for cc sufficiently beyond the characteristic scale Δ−1\Delta^{-1}, the normalized tail satisfies

e−1/Δ∑j>c1j!Δj≤e−Ω(1/Δ),(47)e^{-1/\Delta} \sum_{j>c} \frac{1}{j!\Delta^j} \le e^{-\Omega(1/\Delta)}, \tag*{(47)}

and is therefore strongly suppressed, as Ω\Omega means asymptotic lower bound.

Numerical validation

The following numerical examples illustrate the performance of Algorithm 1 on constrained spin systems of increasing size. In all simulations, the operator pool is constructed from the Krylov expansion truncated at order c=1c = 1 as described in § 4. Therefore,

P=(P1,…,PO(n))T.(48)\mathbf{P} = (P_1,\ldots,P_{\mathcal{O}(n)})^{\mathsf{T}}. \tag*{(48)}

The imaginary time evolution is performed using a constant time step δ=10−2\delta= 10^{-2}. Unless explicitly stated otherwise, these parameters are kept fixed throughout all experiments.

Three-spin Ising model

As a first benchmark, we consider the three-spin triangular Ising model. The Hamiltonian contains only pairwise Ising interaction terms, while the optimization is performed under local constraints on the transverse magnetization X=(X1,…,Xm)T\mathbf{X} = (X_1,\ldots,X_m)^{\mathsf{T}}. The resulting constrained optimization problem is

min⁡∣φ⟩∑(u,v)γuv⟨ZuZv⟩s.t.⟨X⟩=b.(49)\begin{aligned} \min_{|\varphi\rangle}\quad& \sum_{(u,v)} \gamma_{uv}\langle Z_u Z_v\rangle\\ \mathrm{s.t.}\quad& \langle\mathbf{X}\rangle= \mathbf{b}. \tag*{(49)} \end{aligned}

Since the constraint operators do not commute with the Hamiltonian, the constrained minimum differs from the unconstrained Ising ground state, providing a simple yet nontrivial test of the CQITE formalism.

The system is initialized in a product state satisfying the prescribed constraint values, and both CQITE and the exact constrained imaginary time evolution (CITE) are propagated using the same imaginary time step. Figure 1a compares the corresponding energy evolution. The CQITE trajectory closely follows the exact CITE dynamics throughout the evolution and converges to the same constrained minimum, confirming that the unitary approximation accurately reproduces the constrained imaginary time flow.

Figure 1. (a) Energy evolution of CQITE and CITE. (b) Constraint dynamics. Figure 1: CQITE applied to the three-spin Ising model with transverse-magnetization constraints. (a) Energy evolution obtained with homogeneous Ising couplings γ=0.2\gamma= 0.2, comparing CQITE with the exact CITE. (b) Constraint preservation ⟨X⟩\langle X\rangle during the evolution for target values b=(.85,.90,.95)b = (.85,.90,.95), showing that the prescribed constraints are accurately preserved throughout the optimization.

The evolution of the constrained expectation values is shown in Figure 1b. Throughout the propagation the expectation values remain close to the prescribed targets, demonstrating that the Lagrange-multiplier formulation successfully confines the evolution to the desired constraint manifold while simultaneously minimizing the energy. These results provide a first validation of the CQITE algorithm on a constrained optimization problem where the constraints compete directly with the Hamiltonian.

Triangular-ladder spin model

As a second benchmark, we consider the six-spin triangular-ladder model shown in Figure 2. Pairwise interactions XuXv+YuYvX_uX_v + Y_uY_v are defined on the edges, and local magnetization constraint Z=(Z1,…,Zm)T\mathbf{Z} = (Z_1,\ldots,Z_m)^{\mathsf{T}}. The constrained optimization problem is

Six-spin triangular-ladder spin model

Figure 2. Six-spin triangular-ladder spin model. Pairwise interactions γuv(Xu⊗Xv+Yu⊗Yv)\gamma_{uv}(X_u \otimes X_v + Y_u \otimes Y_v) are associated with the edges, while local magnetization constraints ZZ are imposed on each vertex.

min⁡∣φ⟩∑(u,v)γuv⟨XuXv+YuYv⟩s.t.⟨Z⟩=b.(50)\begin{aligned} \min_{|\varphi\rangle}\quad& \sum_{(u,v)} \gamma_{uv}\langle X_uX_v + Y_uY_v\rangle\\ \text{s.t.}\quad& \langle\mathbf{Z}\rangle= \mathbf{b}. \tag*{(50)} \end{aligned}

Figure 3a compares the CQITE evolution with the constrained imaginary time evolution (CITE). The two trajectories are nearly indistinguishable throughout the optimization and converge to the same final energy. This agreement indicates that the Hamiltonian-generated operator pool accurately captures the tangent space explored by the constrained dynamics.

Energy evolution of CQITE and CITE; constraint relative error

Figure 3. CQITE applied to the six-spin triangular-ladder model with local magnetization constraints. (a) Energy evolution obtained with homogeneous couplings γ=0.2\gamma= 0.2, comparing CQITE with the exact CITE. (b) Evolution of the constrained expectation values ⟨Z⟩\langle Z\rangle during the optimization for identical target values b=(.70,.75,.80,.85,.90,.95)\mathbf{b} = (.70,.75,.80,.85,.90,.95).

Finally, the constraint violations are shown in Figure 3b. Throughout the evolution, all six local constraints remain satisfied up to errors of order ∼10−3\sim10^{-3}. Small residual deviations from the target constraint are nevertheless observed. Their origin is beyond the scope of the present work, although they do not appear to be associated with the Krylov operator-pool approximation, whose agreement with the full CITE dynamics remains excellent throughout the evolution.

Two-particle fermionic triangle

As a fermionic benchmark, we consider the two-particle triangle graph studied by Penz and van Leeuwen [2] on which we consider two spin-less fermions. The system is given by the Hamiltonian

H=∑p=1,q=13hpqcp†cq(51)H = \sum_{p=1,q=1}^{3} h_{pq} c_p^\dagger c_q \tag*{(51)}

where

h=(2−1−1−12−1−1−12),(52)h=\begin{pmatrix}2&-1&-1\\-1&2&-1\\-1&-1&2\end{pmatrix}, \tag*{(52)}

which is the graph Laplacian of a triangular lattice with 3 sites. We consider the 2-particle sector. The constraints are given by the on-site particle density operators ρ^=(c1†c1,c2†c2,c3†c3)T\hat{\boldsymbol{\rho}}=(c_1^\dagger c_1,c_2^\dagger c_2,c_3^\dagger c_3)^{\mathsf{T}}. The goal is to minimise the expectation value of the Hamiltonian under specified values ρj=⟨φ∣ρ^j∣φ⟩\rho_j=\langle\varphi|\hat{\rho}_j|\varphi\rangle. This defines the so-called Levy constrained search density-functional

F(ρ)=min⁡∣φ⟩ ⟨H⟩F(\boldsymbol{\rho})=\min_{\lvert\varphi\rangle}\ \langle H\rangle
s.t.⟨ρ^⟩=ρ.(53)\mathrm{s.t.}\quad\langle\hat{\boldsymbol{\rho}}\rangle=\boldsymbol{\rho}. \tag*{(53)}

where we minimize over all wave functions giving a pre-scribed density ρ=(ρ1,ρ2,ρ3)T\boldsymbol{\rho}=(\rho_1,\rho_2,\rho_3)^{\mathsf{T}}. The densities lie on the plane ρ1+ρ2+ρ2=2\rho_1+\rho_2+\rho_2=2 in density space on which the conditions 0≤ρj≤10\leq\rho_j\leq1 cut out a triangular density domain for the function F(ρ)F(\boldsymbol{\rho}). In the occupation basis {∣110⟩,∣101⟩,∣011⟩}\{\lvert110\rangle,\lvert101\rangle,\lvert011\rangle\}, the wavefunction is given by

∣φ⟩=c1∣110⟩+c2∣101⟩+c3∣011⟩(54)\lvert\varphi\rangle=c_1\lvert110\rangle+c_2\lvert101\rangle+c_3\lvert011\rangle \tag*{(54)}

the densities are given by

ρ1=∣c1∣2+∣c2∣2ρ2=∣c1∣2+∣c3∣2ρ3=∣c2∣2+∣c3∣2(55)\rho_1 = |c_1|^2 + |c_2|^2 \qquad\rho_2 = |c_1|^2 + |c_3|^2 \qquad\rho_3 = |c_2|^2 + |c_3|^2 \tag*{(55)}

Since we can split off a global phase we can parametrize

c1=1−ρ3 eiϑ1c2=1−ρ2 eiϑ2c3=1−ρ1.(56)c_1 = \sqrt{1-\rho_3}\,e^{i\vartheta_1} \qquad c_2 = \sqrt{1-\rho_2}\,e^{i\vartheta_2} \qquad c_3 = \sqrt{1-\rho_1}. \tag*{(56)}

and after a short calculation we find that

F(ρ)=4+2min⁡ϑ1,ϑ2T(ϑ1,ϑ2)(57)F(\rho) = 4 + 2\min_{\vartheta_1,\vartheta_2} T(\vartheta_1,\vartheta_2) \tag*{(57)}

where

T(ϑ1,ϑ2)=α1cos⁡ϑ1−α2cos⁡ϑ2−α3cos⁡(ϑ1−ϑ2)(58)T(\vartheta_1,\vartheta_2) = \alpha_1\cos\vartheta_1 - \alpha_2\cos\vartheta_2 - \alpha_3\cos(\vartheta_1-\vartheta_2) \tag*{(58)}

where we defined

α1=(1−ρ1)(1−ρ3)α2=(1−ρ1)(1−ρ2)α3=(1−ρ2)(1−ρ3)(59)\alpha_1 = \sqrt{(1-\rho_1)(1-\rho_3)} \qquad\alpha_2 = \sqrt{(1-\rho_1)(1-\rho_2)} \qquad\alpha_3 = \sqrt{(1-\rho_2)(1-\rho_3)} \tag*{(59)}

The minimisation can be carried out analytically [2] and yields the exact expression:

F(ρ)={3∥ρ−ρ0∥2≤163+(2max⁡k1−ρk−∑j=131−ρj)2∥ρ−ρ0∥2>16(60)F(\rho) = \begin{cases} 3 & \lVert\rho-\rho_0\rVert_2 \le\frac{1}{\sqrt{6}} \\ 3 + \left(2\max_k\sqrt{1-\rho_k}-\sum_{j=1}^{3}\sqrt{1-\rho_j}\right)^2 & \lVert\rho-\rho_0\rVert_2 > \frac{1}{\sqrt{6}} \end{cases} \tag*{(60)}

where ρ0=(23,23,23)T\rho_0 = \left(\frac{2}{3},\frac{2}{3},\frac{2}{3}\right)^{\mathrm{T}} is the equidensity for equal occupation on all sites. The densities satisfying ∥ρ−ρ0∥2≤1/6\lVert\rho-\rho_0\rVert_2 \le1/\sqrt{6} are contained in a disc of radius 1/61/\sqrt{6} in the plane ρ1+ρ2+ρ3=2\rho_1+\rho_2+\rho_3=2 centered at ρ0\rho_0. The densities in this disc correspond to degenerate densities of the Hamiltonian in the absence of any external potential, while the remaining densities in the density domain require a non-zero external potential to generate them as a ground state density. The exception is the set of densities on the boundary of the triangular domain which correspond to full occupation ρj=1\rho_j=1 on one of the sites and would require an infinite potential to generate (these are not vv-representable in density-functional nomenclature), with the exception of the points where the degeneracy disc touches the boundary.

Let us now consider the translation of the problem to qubits. If we write the second quantized Hamiltonian in qubit form using a Jordan-Wigner transformation we find

H=41−12(X1X2+Y1Y2+X2X3+Y2Y3+X1Z2X3+Y1Z2Y3).(61)H = 4\mathbb{1} - \frac{1}{2}(X_1X_2+Y_1Y_2+X_2X_3+Y_2Y_3+X_1Z_2X_3+Y_1Z_2Y_3). \tag*{(61)}

Following the second-quantized formulation introduced in § 3.2, the constraint operators are chosen as the local occupation operators in the qubit representation,

C=(Z1,Z2,Z3)T.(62)C = (Z_1,Z_2,Z_3)^{\mathrm{T}}. \tag*{(62)}

The corresponding target satisfies

b1+b2+b3=−1.(63)b_1+b_2+b_3=-1. \tag*{(63)}

where the relations between ρj\rho_j and bj\mathbf{b}_j are given by b=1−2ρ\mathbf{b}=1-2\boldsymbol{\rho}. The initial state is selected within the fixed-density manifold as a non-stationary state of the Penz–van Leeuwen phase parametrization [2].

To visualize the constrained-search landscape we consider the parametrization of Eq. (57) and display the resulting constrained-search energy landscape and the CQITE trajectory in Figure 4.

Figure 4. CQITE evolution on the Penz–van Leeuwen constrained-search energy landscape in the phase coordinates ϑj−ϑj⋆\vartheta_j-\vartheta_{j^\star}. The background shows the constrained energy range. The algorithm trajectory is represented by points coloured according to the imaginary time τ\tau, evolving from the initial state ∣φ(0)⟩\lvert\varphi(0)\rangle to the final state ∣φ(5)⟩\lvert\varphi(5)\rangle. The origin corresponds to the constrained minimum.

The trajectory illustrates the relation between the CQITE dynamics and the geometry of the constrained energy landscape. Starting from a non-stationary point in the low-energy region, the evolution follows a curved path towards the minimum, with its direction appearing to be influenced by the surrounding higher-energy contours. Thus, the phase-space representation provides an intuitive picture of how the constrained imaginary-time dynamics navigate the local structure of the energy surface. This complements the quantitative agreement between CQITE and CITE observed in the previous benchmarks.

We now consider the full constrained-search functional. For each target density ρ\boldsymbol{\rho}, we initialize the CQITE evolution in the two-particle state

∣φ(0)⟩=1−ρ1∣011⟩+1−ρ2∣101⟩+i1−ρ3∣110⟩.(64)|\varphi(0)\rangle= \sqrt{1-\rho_1}|011\rangle+\sqrt{1-\rho_2}|101\rangle+i\sqrt{1-\rho_3}|110\rangle. \tag*{(64)}

The squared modulus of the amplitude of the basis state with site jj empty is 1−ρj1-\rho_j, and we choose the initial state to be complex since the densities in the degeneracy circle can only be reached by complex wavefunctions [2]. The initial state has ⟨ρ^⟩=ρ\langle\hat{\boldsymbol{\rho}}\rangle=\boldsymbol{\rho}, or equivalently ⟨Z⟩=1−2ρ=b\langle\mathbf{Z}\rangle=1-2\boldsymbol{\rho}=\mathbf{b}. It is normalized since ∑jρj=2\sum_j\rho_j=2.

Figure 5(a) compares the analytic reference over the full density domain with the energies obtained by CQITE at 21 target densities. These targets are obtained by contracting a uniform barycentric grid by a factor of 0.9 toward the domain centroid, preserving its triangular arrangement while placing all samples strictly inside the domain. The three outer branches rise away from the central region, where the functional is constant at its minimum, F(ρ)=3F(\boldsymbol{\rho}) = 3.

Figure 5. (a) Constrained energy and CQITE samples. (b) Energy error across the density domain. Exact constrained-search functional on the density domain of the three-site triangle graph, with converged CQITE energies at sampled target densities.

Figure 5(b) shows the signed energy difference ⟨H⟩−F(ρ)\langle H\rangle- F(\boldsymbol{\rho}) at the same targets. The largest absolute energy difference is approximately 1.4×10−41.4 \times10^{-4}, with the largest deviations occurring at sampled densities in the outer regions.

The constraint potential v\mathbf{v} is the negative gradient of F(ρ)F(\boldsymbol{\rho}) [2]:

v=−∇F(ρ)(65)\mathbf{v} = -\nabla F(\boldsymbol{\rho}) \tag*{(65)}

which gives v=0\mathbf{v} = 0 on the degeneracy disc and non-zero outside it. On the boundaries of the triangular domain, one of the density coordinates corresponds to full occupation, i.e. ρj=1\rho_j = 1. We thus note from the terms 1−ρj\sqrt{1-\rho_j} in the expression for F(ρ)F(\boldsymbol{\rho}) that, at these points, the gradient diverges as (1−ρj)−1/2(1-\rho_j)^{-1/2}, making the constraint potential infinite.

Summary and outlook

In this work, we introduced constrained quantum imaginary-time evolution and derived a systematic Krylov hierarchy for constructing its operator pool. At fixed truncation order, the resulting pool grows polynomially with system size. Numerical results on spin and fermionic models show close agreement with constrained imaginary-time evolution and accurate constraint preservation. Several directions merit further study. One is to investigate the Krylov construction for ground-state preparation with unconstrained Hamiltonians. It would also be useful to derive sharper truncation bounds that better account for the accuracy observed in our numerical experiments. Finally, benchmarking CQITE on quantum hardware could help assess the effects of finite sampling, noise, and circuit depth.

Acknowledgments

All numerical experiments presented in this work were implemented and performed using Qiskit. The source code, together with the scripts required to reproduce the results reported in this paper, is publicly available on GitHub. MS acknowledges support by the IRA Programme, project no. FENG.02.01-IP.05-0006/23, financed by the FENG program 2021–2027, Priority FENG.02, Measure FENG.02.01, with the support of the FNP. LCB acknowledges support from the Polish National Science Centre (NCN) through the Sonata Bis (grant no. UMO-2024/54/E/ST2/00316).

Appendix

Constraint-based reference state

Let

C=(C1,…,Cn)T∈{X,Z}n(66)\mathbf{C} = (C_1,\ldots,C_n)^{\mathrm{T}} \in\{X,Z\}^n \tag*{(66)}

We restrict the constraints to local operators, as motivated by physically relevant scenarios in which target constraint values are imposed through local couplings to external fields or controls. Moreover, restricting the local constraints to {X,Z}\{X,Z\} entails no loss of generality, since any set of local Pauli constraints can be brought to this form by a local Clifford transformation, applied consistently to the Hamiltonian, state, and constraint operators.

We introduce the Pauli vector

Y=(Y1,…,Yn),\mathbf{Y} = (Y_1,\ldots,Y_n),

which anticommutes with each constraint. For constraints acting on distinct qubits [Yj,Ck]=0[Y_j,C_k]=0. Starting from

∣0⟩=∣0⟩⊗n,|0\rangle= |0\rangle^{\otimes n},

we define the constraint-dependent reference state

∣C⟩=⨂j=1n{∣0⟩j,Cj=Zj,∣+⟩j,Cj=Xj.|\mathbf{C}\rangle= \bigotimes_{j=1}^{n} \begin{cases} |0\rangle_j, & C_j=Z_j,\\ |+\rangle_j, & C_j=X_j. \end{cases}

Equivalently,

∣C⟩=e+iπ4(1−⟨0∣C∣0⟩)TY∣0⟩.|\mathbf{C}\rangle= e^{+\mathrm{i}\frac{\pi}{4}(1-\langle0|\mathbf{C}|0\rangle)^{\mathrm{T}}\mathbf{Y}}|0\rangle.

The target constraint values b\mathbf{b} are then imposed by the rotation

∣φ(0)⟩=e−iarccos⁡b2Y∣C⟩.|\varphi(0)\rangle= e^{-\mathrm{i}\frac{\arccos\mathbf{b}}{2}\mathbf{Y}}|\mathbf{C}\rangle.

By construction,

⟨φ(0)∣C∣φ(0)⟩=b.\langle\varphi(0)|\mathbf{C}|\varphi(0)\rangle=\mathbf{b}.

References

  1. [1]M. Penz and R. van Leeuwen, “Constrained Search in Imaginary Time,” Physical Review A 112, 032815 (2025).DOI
  2. [2]M. Penz and R. van Leeuwen, “Density-functional theory on graphs,” Journal of Chemical Physics 155, 244111 (2021).DOI
  3. [3]M. Motta, C. Sun, A. T. K. Tan, et al., “Determining eigenstates and thermal states on a quantum computer using quantum imaginary time evolution,” Nature Physics 16, 205–210 (2020).DOI
  4. [5]S. McArdle, T. Jones, S. Endo, Y. Li, S. C. Benjamin, and X. Yuan, “Variational ansatz-based quantum simulation of imaginary time evolution,” npj Quantum Information 5, 75 (2019).DOI
  5. [6]X. Yuan, S. Endo, Q. Zhao, Y. Li, and S. C. Benjamin, “Theory of variational quantum simulation,” Quantum 3, 191 (2019).DOI
  6. [7]S. Endo, J. Sun, Y. Li, S. C. Benjamin, and X. Yuan, “Variational quantum simulation of general processes,” Physical Review Letters 125, 010501 (2020).DOI
  7. [8]H. R. Grimsley, S. E. Economou, E. Barnes, and N. J. Mayhall, “An adaptive variational algorithm for exact molecular simulations on a quantum computer,” Nature Communications 10, 3007 (2019).DOI
  8. [9]Y. S. Yordanov, V. Armaos, C. H. Baldwin, I. G. Ryabinkin, T.-C. Yen, S. Genin, and A. F. Izmaylov, “Qubit-ADAPT-VQE: An adaptive algorithm for constructing hardware-efficient ansatze on a quantum processor,” Nature Communications 12, 2260 (2021).
  9. [10]M. Levy, “Universal variational functionals of electron densities, first-order density matrices, and natural spin-orbitals and solution of the v-representability problem,” Proceedings of the National Academy of Sciences USA 76, 6062–6065 (1979).pubmed.ncbi.nlm.nih.gov/16592733
  10. [11]E. H. Lieb, “Density functionals for Coulomb systems,” International Journal of Quantum Chemistry 24, 243–277 (1983).
  11. [12]C. Lanczos, “An iteration method for the solution of the eigenvalue problem of linear differential and integral operators,” Journal of Research of the National Bureau of Standards 45, 255–282 (1950).DOI
  12. [13]T. J. Park and J. C. Light, “Unitary quantum time evolution by iterative Lanczos reduction,” Journal of Chemical Physics 85, 5870–5876 (1986).DOI
  13. [15]C. L. Cortes and S. K. Gray, “Quantum Krylov subspace algorithms for ground- and excited-state energy estimation,” Physical Review A 105, 022417 (2022).DOI
  14. [16]N. H. Stair, R. Huang, and F. A. Evangelista, “A multireference quantum Krylov algorithm for strongly correlated electrons,” Journal of Chemical Theory and Computation 16, 2236–2245 (2020).
  15. [17]W. Kirby, M. Motta, and A. Mezzacapo, “Exact and efficient Lanczos method on a quantum computer,” Quantum 7, 1018 (2023).DOI
  16. [18]Z. Zhang, A. Wang, X. Xu, and Y. Li, “Measurement-efficient quantum Krylov subspace diagonalisation,” Quantum 8, 1438 (2024).DOI
  17. [19]W. Kirby et al., “Observation of Improved Accuracy over Classical Sparse Ground-State Solvers using a Quantum Computer,” arXiv preprint arXiv:2603.03496, 2026.arxiv.org/abs/2603.03496
  18. [20]J. Li and S. C. Benjamin, “Efficient variational quantum simulator incorporating active error minimization,” Physical Review X 7, 021050 (2017).DOI
  19. [23]A. D. McLachlan, “A variational solution of the time-dependent Schrödinger equation,” Molecular Physics 8, 39–44 (1964).DOI
  20. [24]M. Gluza, J. Son, B. H. Tiang, R. Zander, R. Seidel, Y. Suzuki, Z. Holmes, and N. H. Y. Ng, “Double-Bracket Quantum Algorithms for Quantum Imaginary-Time Evolution,” Physical Review Letters 136, 020601 (2026).arxiv.org/abs/2412.04554
  21. [25]D. Cuomo and R. van Leeuwen, “Quantum simulation in imaginary time for gauge-invariant models,” Physica Scripta 101, 035110 (2026).
  22. [26]Y. Mao, M. Chaudhary, M. Kondappan, J. Shi, E. O. Ilo-Okeke, V. Ivannikov, and T. Byrnes, “Measurement-Based Deterministic Imaginary Time Evolution,” Phys. Rev. Lett. 131, 110602 (2023).arxiv.org/abs/2202.09100

Paper details

vlnc:W_n8avwa93
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 15

downloads · 1

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