Quantum Inverse Iteration

• Quantum Inverse Iteration is a spectral transformation technique that adapts classical inverse power iteration to quantum systems, amplifying target eigenstates via realizable quantum primitives.
• It employs methods like Fourier decomposition, continuous-variable integrals, and block-encoding to approximate non-unitary inverses using native quantum operations such as Hamiltonian evolution.
• Diverse formulations, including Q-Inv, I-Iter, and EF-based QIPI, offer varied trade-offs in accuracy, resource cost, and convergence for both ground and excited state preparation.

Quantum inverse iteration is a family of spectral-transformation algorithms that adapts classical inverse power iteration to quantum settings by replacing repeated applications of $(H-\omega I)^{-1}$ or $H^{-k}$ with realizable quantum primitives. In its basic form, the method starts from a trial state with nonzero overlap with the target eigenstate and amplifies that component through the inverse of a shifted Hamiltonian, so that the target eigenvalue becomes dominant under iteration. In the literature, this idea appears in several technically distinct forms: Fourier or Laplace-like decompositions into coherent time evolutions, continuous-variable realizations of inverse maps, block-encoding and generalized quantum signal processing constructions, and variational or projected inverse-iteration schemes that preserve the same spectral logic without directly implementing a non-unitary inverse on hardware (Kyriienko, 2019, Khinevich et al., 15 Jul 2025, Patil et al., 26 Jun 2026).

1. Spectral principle and mathematical basis

The classical prototype is shifted inverse iteration,

$$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$

which converges to the eigenvector whose eigenvalue is closest to the shift $\tau$. In the unshifted positive-definite case one can write

$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$

so that excited-state components are suppressed relative to the ground-state component, and the energy is estimated by the Rayleigh quotient

$$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$

This is the conceptual basis of the “quantum inverse iteration algorithm for programmable quantum simulators” and of later Q-Inv formulations for quantum chemistry Hamiltonians (Kyriienko, 2019, Cainelli et al., 2024).

A central distinction in the quantum literature is between direct and indirect realizations of the inverse. In “Q‑Inv,” powers of the inverse are implemented directly from an integral representation of $H^{-k}$, whereas in “I‑Iter” one first approximates $H^{-1}$ and then applies that approximation $k$ times. In the direct formulation, the same set of unitary propagators $e^{-i\lambda_j H}$ can be reused for all $H^{-k}$0, with only the coefficients changing; in the iterative formulation, the number of inverse applications grows as $H^{-k}$1 (Cainelli et al., 2024). This separation between spectral filtering and implementation mechanism has become a defining feature of the field.

A second distinction is between inverse maps aimed primarily at the ground state and shifted inverse maps aimed at arbitrary interior eigenvalues. Earlier hardware-oriented work focused on ground-state preparation with positive or shifted Hamiltonians. More recent QSVT-based formulations treat

$H^{-k}$2

as the core filter for arbitrary excited states, provided $H^{-k}$3 is chosen so that the target eigenvalue is uniquely closest to the shift (Patil et al., 26 Jun 2026).

2. Fourier, quadrature, and time-evolution realizations

The most direct hardware-native construction expresses inverse powers as integrals over unitary time evolution. A representative identity used in Q‑Inv is

$H^{-k}$4

which reduces the non-unitary inverse power to a weighted superposition of propagators $H^{-k}$5 with $H^{-k}$6. After discretization, this becomes a linear combination of unitaries, and on programmable quantum devices the algorithm is reformulated as separate measurements of overlaps between the initial and propagated wavefunctions (Kyriienko, 2019, Cainelli et al., 2024).

The 2019 programmable-simulator formulation benchmarked this strategy on molecular hydrogen and beryllium hydride, and also on a Bose–Hubbard model. It used Hartree–Fock product states for molecules and a Mott product state for the Bose–Hubbard example, and reconstructed the action of $H^{-k}$7 from overlap measurements rather than by explicitly preparing the filtered state. The same framework also supports other observables once the inverse-filtered overlap structure is known (Kyriienko, 2019).

A continuous-variable variant replaces discrete linear combinations by a qumode-mediated integral. There the joint unitary

$H^{-k}$8

acts on the system and a continuous-variable ancilla, and the postselected matrix element

$H^{-k}$9

implements the ideal inverse in the infinite-squeezing limit. With finite squeezing $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$0, the realized map approximates $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$1 with state error $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$2, while the success probability scales as $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$3. The same work also analyzed a hybrid qubit-only version that directly sums Hamiltonian evolutions with different durations, concluding that continuous-variable resources are valuable for reducing the coherent evolution time of Hamiltonians in quantum algorithms (He et al., 2020).

These realizations share a common advantage and a common cost. The advantage is that they use primitives natural on quantum hardware—Hamiltonian evolution, overlap measurement, and postselection—without requiring explicit phase estimation. The cost is that non-unitarity is recovered only after numerical quadrature or postselection, so accuracy depends sensitively on discretization, cutoffs, squeezing, or overlap-estimation noise.

3. Polynomial, block-encoding, and filtering formulations

A separate line of work replaces time-evolution integrals by polynomial spectral transformations implemented through block-encoding, QSVT, or generalized quantum signal processing. In the GQSP unification of quantum power methods, quantum inverse iteration is the shifted-inverse member of a broader family that also includes quantum power iteration, power Lanczos, and folded-spectrum methods. There the target map is

$$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$4

and the inverse is approximated by the Taylor expansion

$$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$5

which is then translated into a GQSP-compatible polynomial in the block-encoded Hamiltonian. This eliminates Suzuki–Trotter decomposition, gives a query count equal to the polynomial degree, and was reported to outperform existing inverse iteration variants based on time-evolution operators on the molecular benchmarks considered (Khinevich et al., 15 Jul 2025).

The main limitation of that Taylor-based formulation is its convergence domain: it requires $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$6. The paper explicitly notes that this is natural for ground states with $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$7 slightly above the lowest eigenvalue, but restrictive for excited states. That restriction motivated a QSVT-based quantum inverse power iteration specialized to arbitrary excited-state targeting through filtering polynomials (Khinevich et al., 15 Jul 2025, Patil et al., 26 Jun 2026).

In the excited-state QIPI formulation, one shifts and scales the Hamiltonian,

$$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$8

so that the target eigenvalue is mapped close to zero. Instead of approximating $$|\psi_{k+1}\rangle \propto (H-\tau I)^{-1} |\psi_k\rangle,$$9 directly, the algorithm uses the eigenstate filtering polynomial

$\tau$0

which satisfies $\tau$1 and

$\tau$2

Because the polynomial is symmetric about the origin and sharply peaked near zero, EF-based QIPI was found to be substantially more robust than Chebyshev inverse approximations and other decomposition-based approaches, avoiding divergence with respect to $\tau$3 and efficiently suppressing off-target eigenstates even in closely spaced spectra. Assuming standard oracle access to the Hamiltonian, the paper gives query complexity $\tau$4 with amplitude amplification, where $\tau$5 is the initial target overlap and $\tau$6 the desired accuracy (Patil et al., 26 Jun 2026).

Formulation	Core transform	Characteristic feature
Fourier / Q‑Inv	$\tau$7	Same unitary grid can be reused for all $\tau$8
Continuous-variable QuIPI	$\tau$9	Reduces coherent Hamiltonian evolution time
GQSP/QSVT inverse iteration	$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$0	Avoids Suzuki–Trotter decomposition
EF-based QIPI	$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$1	Robust arbitrary excited-state targeting

4. Numerical behavior, shifts, and stability

The most detailed numerical analysis of Q‑Inv on quantum chemistry Hamiltonians studied H$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$2, LiH, BeH$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$3, and the strongly correlated square H$|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$4 geometry. It found that Q‑Inv gives lower energy results than I‑Iter up to a certain $|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$5, after which the energy increases because the numerical integration error begins to dominate. The dependence on the integration interval is central: for fixed cutoffs there is a maximum effective $|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$6, beyond which the discretized inverse no longer tracks the ideal $|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$7 accurately (Cainelli et al., 2024).

That study also compared trapezoidal quadrature, pure Gauss–Legendre quadrature, and a hybrid scheme. The main empirical conclusion was that Gauss–Legendre quadrature is highly effective for the $|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$8-integration, often with $|\Psi_k\rangle \propto H^{-k}|\Psi_0\rangle,$9, whereas trapezoidal integration works better for the oscillatory $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$0-dependence. The recommended procedure for unknown systems was therefore Gauss–Legendre in $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$1 with $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$2, trapezoidal in $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$3, an initial $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$4 cutoff, tuning the $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$5-cutoff at $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$6, and then increasing $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$7 until the energy reaches a minimum or plateau. For difficult cases such as H$$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$8, the paper proposed a hybrid Q‑Inv + I‑Iter refinement, which uses Q‑Inv up to the best $$E_k = \frac{\langle \Psi_k|H|\Psi_k\rangle}{\langle \Psi_k|\Psi_k\rangle}.$$9 and then applies a small number of inverse-iteration steps (Cainelli et al., 2024).

Earlier programmable-simulator benchmarks exhibited the same tradeoff from a hardware perspective. For H$H^{-k}$0, ideal inverse iteration reached chemical precision at $H^{-k}$1; for BeH$H^{-k}$2, ideal inverse iteration reached chemical precision at $H^{-k}$3. In the Fourier approximation, larger $H^{-k}$4 improved fidelity to the inverse but increased the cost of the propagators. The paper also reported dephasing-noise simulations and error-mitigation via zero-noise extrapolation, showing that chemically accurate estimates remained accessible on small systems for moderate iteration counts (Kyriienko, 2019).

In the GQSP setting, the numerical picture is different. For ground states, high-degree QII polynomials can reach chemical accuracy in very few iterations, but low-degree truncations exhibit early saturation. For excited states, Cheb-inv can display pseudo-convergence, because the truncated approximation to $H^{-k}$5 develops peaks away from the origin; EF-based QIPI avoids this pathology precisely because the filter is constructed to be maximal at the shifted target and uniformly small outside the chosen window (Khinevich et al., 15 Jul 2025, Patil et al., 26 Jun 2026).

5. Alternative inverse primitives and broader algorithmic connections

Several papers develop inverse maps that are not themselves inverse-iteration eigensolvers but supply primitives directly relevant to quantum inverse iteration. One example is matrix inversion by continuous-time quantum walk. There a sparse matrix $H^{-k}$6 is embedded into a larger Hamiltonian, weakly coupled to ancilla levels, and evolved for time $H^{-k}$7. In the base 4-block construction, the postselected output amplitudes satisfy

$H^{-k}$8

while a 6-block construction improves this to

$H^{-k}$9

The authors explicitly note that replacing $H^{-1}$0 by $H^{-1}$1 would make the same construction implement $H^{-1}$2, which is precisely the shift-and-invert primitive required by inverse iteration (Kay et al., 8 Aug 2025).

A different connection comes from block-encoding-based “fast inversion.” For diagonal, normal, and certain 1-sparse matrices, the inverse can be block-encoded directly through reversible classical arithmetic on eigenvalues, so the number of queries to the defining oracles is independent of the condition number at the block-encoding stage. This primitive was then used as a preconditioner in quantum linear-system solvers and in algorithms for Green’s functions and Gibbs-state preparation. The same paper emphasizes that total state-preparation complexity must still depend on

$H^{-1}$3

and that the $H^{-1}$4 dependence is optimal (Tong et al., 2020).

There is also a purely algebraic perspective. For a finite-dimensional matrix $H^{-1}$5, the confluent Vandermonde formalism gives an exact degree-$H^{-1}$6 polynomial representation of any analytic matrix function $H^{-1}$7, including $H^{-1}$8 and $H^{-1}$9, with coefficients determined solely by eigenvalues and multiplicities. This provides a closed-form finite-dimensional specification of the exact target polynomial for inverse filters, although the same paper stresses that full spectral knowledge and Vandermonde conditioning make it primarily a theoretical and small-system tool rather than a scalable inverse-iteration implementation (Hedemann, 2017).

At the opposite end of the spectrum, explicit determinant-and-cofactor circuits can construct a quantum state encoding $k$0 and then solve $k$1 by quantum matrix multiplication. Their depth scales as $k$2, but the success probability is $k$3, so they are not efficient in the usual complexity-theoretic sense. They nevertheless show that explicit inverse maps can be realized without HHL, QSVT, or Hamiltonian simulation (Zenchuk et al., 2024).

6. Variational, PDE, and nonlinear generalizations

The inverse-iteration paradigm has also migrated beyond direct state preparation on quantum hardware. In neural quantum states, Projected Inverse Iteration reformulates ground-state optimization as a projected eigenvalue problem on the variational tangent space. Defining

$k$4

PII replaces stochastic reconfiguration by the linear system

$k$5

The paper interprets this as a natural-gradient eigensolver tailored to inverse iteration, with gap-insensitive convergence in the underlying non-parametric analysis and strong numerical performance on two-dimensional spin models including the frustrated $k$6-$k$7 model (Zhang et al., 5 Jun 2026).

For Schrödinger eigenvalue problems in continuous space, an inverse-operator evolution model replaces explicit orthogonalization by a quasi-orthogonal flow

$k$8

posed on $k$9. The associated discrete scheme combines a Cayley step with an inverse step, requires no explicit orthogonalization, and admits monotonic energy decay, exponential convergence, and time-step bounds independent of the spatial mesh (Wang et al., 24 May 2026). This is not a quantum-computing algorithm in the circuit sense, but it is an inverse-iteration analogue at PDE level.

In nonlinear quantum models the same idea survives through linearization. For spin-orbit coupled Bose–Einstein condensates, the J-method defines a nonlinear inverse iteration

$e^{-i\lambda_j H}$0

where $e^{-i\lambda_j H}$1 is the shifted Jacobian of the nonlinear eigenproblem. The paper proves local linear convergence with factor essentially

$e^{-i\lambda_j H}$2

and reports superlinear convergence numerically when adaptive shifts are used (Henning et al., 13 Jan 2026).

7. Limitations and current research directions

The literature makes clear that “quantum inverse iteration” is not a single algorithmic object but a design pattern. Fourier-based Q‑Inv methods are natural on programmable simulators, but they are highly sensitive to cutoffs, quadrature choices, and iteration order, and they can converge to the wrong state when the effective filter is distorted, as the square H$e^{-i\lambda_j H}$3 example demonstrates (Cainelli et al., 2024). Taylor-based GQSP inverse iteration avoids time evolution and has transparent query counting, but its convergence domain $e^{-i\lambda_j H}$4 favors ground states and motivates alternative filters for interior eigenvalues (Khinevich et al., 15 Jul 2025). Continuous-variable and postselected formulations reduce coherent evolution depth or ancilla cost, but success probability, squeezing, and amplitude amplification remain central resource bottlenecks (He et al., 2020).

Recent filtering-polynomial work suggests a more stable route for excited states. EF-based QIPI was developed precisely because prior quantum inverse power approaches based on Fourier decompositions of the inverse Hamiltonian were highly sensitive to hyperparameter choices and had been observed to be numerically unstable, effectively restricting their use to ground-state preparation (Patil et al., 26 Jun 2026). This suggests that future progress will likely continue to move away from direct quadrature of $e^{-i\lambda_j H}$5 toward structured filters, preconditioned resolvents, and problem-adapted inverse transforms.

Across these variants, the shared theme is unchanged: inverse iteration remains a spectral amplifier. What changes from one formulation to another is the realization of the inverse map—time evolution, qumodes, weakly coupled quantum walks, block-encoded polynomials, projected tangent-space solvers, or nonlinear Jacobian inverses—and the numerical or physical regime in which that realization is stable.