Deterministic MITE for Quantum Imaginary-Time Evolution

• MITE is a quantum algorithm that uses measurement feedback to deterministically simulate imaginary-time evolution without relying on postselection.
• It replaces nonunitary dynamics of e^(–τH) with adaptive unitary corrections, effectively suppressing excited-state amplitudes to target ground states.
• Demonstrations in AKLT state preparation, QND-based filtering, and lattice gauge theory highlight both its efficiency gains and challenges in scalability.

Measurement-based deterministic imaginary time evolution (MITE) denotes a class of quantum algorithms that use measurements together with adaptive or compiled unitary corrections to realize the state-filtering effect of imaginary-time evolution without relying on postselection as the primary convergence mechanism. The common objective is the standard imaginary-time projection

$\ket{\psi(\tau)}=e^{-\hat H \tau}\ket{\psi(0)},$

or its normalized variant, so that excited-state amplitudes are suppressed relative to the target eigenspace. In the literature represented here, MITE appears in two closely related forms: direct measurement-and-feedback protocols, where weak measurements steer the system toward a chosen energy sector and feedback removes undesired branches, and deterministic QITE constructions, where each short nonunitary step is replaced by a measured effective unitary obtained from local correlators (Chen et al., 2023, Kondappan et al., 2022, Sekiyama et al., 20 Apr 2026).

1. Conceptual basis and scope

Imaginary-time evolution is a nonunitary transformation used for ground-state preparation, eigenstate filtering, and related projection tasks. For a time-independent Hamiltonian $H$, if the initial state has nonzero overlap with the ground state, repeated application of $e^{-\tau H}$ suppresses higher-energy components and drives the state toward the ground sector. The central obstacle on quantum hardware is that $e^{-\tau H}$ is nonunitary, whereas native circuit evolution is unitary. MITE addresses this mismatch by using measurement-induced filtering or measured local reconstructions so that the effective update reproduces imaginary-time dynamics at the state level rather than merely at the level of an optimized classical objective (Turro et al., 2021).

The term “deterministic” is used in a specific operational sense. In the direct MITE formulations, measurement outcomes remain stochastic, but convergence does not depend on discarding unfavorable runs. Instead, outcome-dependent control is used so that the target eigenspace is the only stable fixed point of the closed-loop evolution (Kondappan et al., 2022). In deterministic QITE formulations, each short imaginary-time factor is replaced by a unitary generated by a measured effective Hermitian operator, so the algorithm proceeds by repeated measured reconstruction and compiled unitary application rather than heralded success/failure branches (Sekiyama et al., 20 Apr 2026).

This distinguishes MITE from ancilla-postselection approaches. In heralded schemes, a unitary dilation of a nonunitary step is followed by ancilla measurement, and only the favorable branch realizes the intended update. Repeated steps then suffer from decaying total success probability, which is precisely the failure mode that deterministic measurement-based constructions aim to avoid (Turro et al., 2021).

2. Local measurement-and-feedback update rules

A concrete MITE update appears in the AKLT-state preparation protocol. There, MITE is described as using “a combination of quantum measurements with adaptive unitary transformations to deterministically prepare an eigenstate of a given Hamiltonian.” The target is the 1D AKLT ground state, which satisfies

$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$

with $\hat{\mathcal{P}}_{j,j+1}^{S=2}$ the local spin-2 bond projector. The protocol therefore replaces global filtering by repeated local filtering onto the kernel of each bond projector (Chen et al., 2023).

The local weak-measurement operators are

$$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$

with $q\in\{0,1\}$ and local eigenvalues $E_{S,m}=\delta_{S,2}$. Operationally, this measurement is implemented by preparing an ancilla in $\ket{+}_a$, applying

$H$0

and measuring the ancilla in the $H$1-basis. After outcome $H$2, the normalized state update is

$H$3

The measurement record is compressed into counts $H$4, from which the protocol estimates the peak location

$H$5

If this estimate remains below a threshold $H$6, the feedback unitary is the identity. Otherwise a corrective unitary $H$7 is applied and the counters are reset. In the AKLT construction, $H$8 is chosen as a random local spin rotation on the two sites. The whole chain is updated by alternating parallel odd-bond and even-bond sweeps, so the local MITE subroutines on disjoint bonds commute (Chen et al., 2023).

This architecture captures the core logic of direct MITE: the measurement performs a weak imaginary-time-like spectral filter, while adaptive control prevents the trajectory from stabilizing on an undesired branch.

3. Quantum nondemolition realization and measurement nonlinearities

A more general measurement-based deterministic imaginary-time construction replaces the earlier weak energy-basis measurement gadget by quantum nondemolition (QND) measurements of the Hamiltonian itself. The QND interaction is

$H$9

where $e^{-\tau H}$0 is the optical mode coupled to the system and $e^{-\tau H}$1 is the dimensionless interaction time. If

$e^{-\tau H}$2

then photon-count outcomes $e^{-\tau H}$3 after the interferometric readout produce the conditioned state

$e^{-\tau H}$4

with measurement operator

$e^{-\tau H}$5

Repeated QND measurements therefore generate a cumulative energy-diagonal filter whose peak position can be estimated from the measurement record (Kondappan et al., 2022).

In the regime $e^{-\tau H}$6, the cumulative counts define an energy estimator

$e^{-\tau H}$7

The closed-loop state update is written as

$e^{-\tau H}$8

where $e^{-\tau H}$9 if the estimated energy lies within the target window and $e^{-\tau H}$0 otherwise. In the long-time regime, an additional phase correction $e^{-\tau H}$1 removes the sign factor generated by the QND response. The protocol is deterministic in the sense that every run is continued with feedback; no postselection is required (Kondappan et al., 2022).

The same formalism yields an explicit route from collective single-qubit QND couplings to effective many-body imaginary-time operators. For short interaction time, a collective Hamiltonian such as

$e^{-\tau H}$2

produces a measurement operator containing $e^{-\tau H}$3, and hence a genuine interaction term proportional to $e^{-\tau H}$4. For long interaction time $e^{-\tau H}$5, the identity

$e^{-\tau H}$6

converts a sum of single-qubit observables into an effective $e^{-\tau H}$7-body operator inside the measurement filter. This mechanism is used to prepare a four-qubit cluster state using only collective single-qubit QND measurements and single-qubit adaptive operations (Kondappan et al., 2022).

4. Deterministic QITE as a measurement-based implementation

A second major realization of MITE is deterministic QITE in the Motta-style sense: each local normalized imaginary-time step is approximated by a unitary generated by a measured effective Hermitian operator. For a Hamiltonian decomposed as

$e^{-\tau H}$8

the target local step is

$e^{-\tau H}$9

and the deterministic approximation is

$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$0

with

$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$1

The coefficients are obtained from the linear system

$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$2

where

$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$3

All expectation values are taken in the current normalized state, so a hardware realization would measure these correlators, solve the linear system classically, and apply the resulting effective unitary. In that algorithmic sense, deterministic QITE is a measurement-based deterministic imaginary-time method (Sekiyama et al., 20 Apr 2026).

For two-dimensional pure $$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$4 lattice gauge theory, the principal advance is symmetry restriction of the Pauli pool. The full Pauli pool on a support $$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$5 is reduced in three exact stages: keep only strings with an odd number of $$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$6's, keep only strings commuting with all Gauss operators on the support, and optionally quotient by Gauss-law equivalence classes. The gauge-invariant condition is expressed by requiring the $$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$7-support of the Pauli string to form closed loops. This reduction leaves the deterministic update unchanged while sharply reducing measurement and gate costs (Sekiyama et al., 20 Apr 2026).

Support	Full nontrivial pool	Reduced gauge-invariant odd-$$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$8 pool
One plaquette	255	8
Two adjacent plaquettes	16383	192
Three in a row	1048575	2048

The measurement burden scales as $$\hat{\mathcal{P}}_{j,j+1}^{S=2}\ket{\mathrm{AKLT}}=0 \qquad \forall j,$$9 to populate the linear system, and implementing the effective unitary scales as $\hat{\mathcal{P}}_{j,j+1}^{S=2}$0 in the worst case. This makes symmetry reduction structurally central rather than incidental (Sekiyama et al., 20 Apr 2026).

5. Efficient specialized realization: AKLT state preparation

The AKLT construction is the clearest example of a model-specific MITE speedup. The AKLT Hamiltonian is

$\hat{\mathcal{P}}_{j,j+1}^{S=2}$1

and the periodic-chain AKLT state is the unique zero-energy ground state satisfying

$\hat{\mathcal{P}}_{j,j+1}^{S=2}$2

Because the target is a common zero-eigenvalue state of all local bond projectors, the protocol can use repeated local MITE steps on two-site subspaces of fixed dimension $\hat{\mathcal{P}}_{j,j+1}^{S=2}$3 instead of confronting the full Hilbert-space dimension $\hat{\mathcal{P}}_{j,j+1}^{S=2}$4. The paper argues that this yields approximately constant preparation time with chain length and presents numerical evidence supporting that conclusion (Chen et al., 2023).

Two diagnostics are used. The local one is the partial bond fidelity

$\hat{\mathcal{P}}_{j,j+1}^{S=2}$5

and the global one is

$\hat{\mathcal{P}}_{j,j+1}^{S=2}$6

Averaged over 100 runs, the partial fidelities were reported to converge to unity typically within $\hat{\mathcal{P}}_{j,j+1}^{S=2}$7 measurement steps in each local MITE subroutine. The total fidelity reached near unity within about $\hat{\mathcal{P}}_{j,j+1}^{S=2}$8 odd/even rounds, with little dependence on the tested system size. A simplified direct-projection study,

$\hat{\mathcal{P}}_{j,j+1}^{S=2}$9

gave an approximately constant round count $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$0 for $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$1 to $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$2, supporting the same interpretation (Chen et al., 2023).

The paper also analyzes hardware compatibility. Each spin-1 site is embedded into two qubits via

$$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$3

and the local measurement Hamiltonian is

$$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$4

A variational circuit recompilation of the corresponding five-qubit unitary produced local circuits with maximum operator fidelity above $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$5 for layer depth $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$6, and the recompiled circuit used on the order of $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$7 CNOT gates, compared with 499 CNOT gates from default Qiskit transpilation (Chen et al., 2023).

The significance of the AKLT result is therefore not a universal MITE complexity theorem, but a sharp demonstration that frustration-free projector structure can make deterministic MITE dramatically more efficient than naive Hilbert-space estimates suggest.

6. Contrast with probabilistic and hybrid alternatives

The boundary of MITE is clarified by comparison with probabilistic ancilla-postselection schemes. In the ancilla-assisted imaginary-time propagation proposal on a quantum chip, the system state is extended by one ancilla and evolved under a block-encoded unitary $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$8; the desired imaginary-time map is obtained only when the ancilla is measured in $$M_{j,j+1}^q= \frac{1}{2}\sum_{S=0}^2 \sum_{m=-S}^{S} \left(\cos \epsilon E_{S,m}-(-1)^q \sin\epsilon E_{S,m}\right) |S,m\rangle_{j,j+1}\langle S,m|_{j,j+1},$$9. For repeated short-time evolution, the success probability obeys

$q\in\{0,1\}$0

and in the favorable regime this is bounded below by $q\in\{0,1\}$1, so total success decays exponentially with the number of steps. This is explicitly a heralded, postselected, nonvariational method rather than deterministic MITE (Turro et al., 2021).

A similarly direct contrast appears in non-unitary Trotter circuits for imaginary time evolution. There, the standard Pauli gadget is modified by replacing its central $q\in\{0,1\}$2 rotation with an ancilla-mediated nonunitary primitive. The resulting global success probability after $q\in\{0,1\}$3 Trotter steps satisfies

$q\in\{0,1\}$4

with $q\in\{0,1\}$5 the Hamiltonian 1-norm, so the method is explicitly probabilistic even though it is measurement-based and ancilla-assisted (Leadbeater et al., 2023).

A partial bridge between heralded and deterministic regimes is provided by probabilistic imaginary-time evolution combined with quantum amplitude amplification. In that framework, a tunable parameter $q\in\{0,1\}$6 can be optimized as

$q\in\{0,1\}$7

so that the amplified success probability becomes unity for a chosen amplification count $q\in\{0,1\}$8. The resulting ancilla-separable output is called deterministic imaginary-time evolution in that paper, but it is derived from a postselection-originated ancilla construction rather than from measurement-feedback branch unification (Nishi et al., 2022).

Distinct neighboring families remain measurement-driven without being MITE in the narrow sense. Operator-Projected Variational Quantum Imaginary Time Evolution enforces imaginary-time dynamics only on a chosen operator set and replaces full-state metric estimation by an observable-space projected linear system, reducing measurement complexity from quadratic to linear in the number of parameters and avoiding fidelity-estimation circuits (Anuar et al., 2024). ITEMC for QUBO optimization likewise uses measured one- and two-qubit moments to fit local unitary gates that mimic commuting imaginary-time factors, but it is a hybrid deterministic surrogate rather than a direct deterministic physical realization of the nonunitary step (Chai et al., 28 May 2025).

7. Performance envelope, limitations, and current frontier

The most extensive accuracy benchmark in the present set of papers is deterministic QITE for two-dimensional pure $q\in\{0,1\}$9 lattice gauge theory. Using symmetry-restricted Pauli pools and tensor-network simulation, the method reached relative ground-state energy error below $E_{S,m}=\delta_{S,2}$0 for systems up to twelve plaquettes and couplings $E_{S,m}=\delta_{S,2}$1 in the regime studied. For the smallest $E_{S,m}=\delta_{S,2}$2 lattice at $E_{S,m}=\delta_{S,2}$3, both ITE and QITE errors dropped below $E_{S,m}=\delta_{S,2}$4 by $E_{S,m}=\delta_{S,2}$5. At stronger coupling, however, QITE error saturated as $E_{S,m}=\delta_{S,2}$6, which the paper interprets as finite Pauli-pool and finite-domain error rather than Trotter error. The discrepancy also grew with system size, indicating that locality truncation becomes the dominant approximation when correlations strengthen (Sekiyama et al., 20 Apr 2026).

The AKLT study identifies a different limitation profile. Its approximately constant scaling with chain length is explicitly presented as a special case that relies on the AKLT state being a common zero eigenstate of all bond projectors. The paper states that no rigorous proof of constant-time scaling is given, that the exceptional efficiency is model-specific, and that it remains unresolved which other models admit comparable behavior. Its noise study is limited to simple random local rotation models, and the full algorithm was not demonstrated on hardware (Chen et al., 2023).

These limits delineate the current frontier of MITE. Exact symmetry restrictions can remove measurement overhead without changing the deterministic update, as in lattice gauge theory, but finite local domains still cap accuracy (Sekiyama et al., 20 Apr 2026). Frustration-free projector structure can produce striking efficiency gains, but those gains do not yet generalize beyond specially structured targets such as AKLT (Chen et al., 2023). At the same time, adjacent long-time coherent QITE methods based on adaptive normalization and QPP show that the normalization bottleneck itself can sometimes be stabilized, although those methods remain heralded rather than deterministic in the strict MITE sense (Zhang et al., 1 Jul 2025).

Across these developments, the defining technical distinction remains stable. MITE is not merely any quantum algorithm inspired by imaginary time, nor any measurement-assisted nonunitary simulation. It denotes constructions in which measurement records are used to realize or reconstruct the imaginary-time update without making convergence contingent on exponentially rare success branches. The literature so far shows that this goal is achievable in several non-equivalent ways—adaptive weak-measurement feedback, QND-based filtering, and deterministic QITE via measured local generators—but also that scalability is presently tied to locality, symmetry, spectral structure, and the availability of problem-specific operator reductions.