Time Rewinding: Concepts & Protocols

• Time rewinding is a family of methods that reconstruct, emulate, or infer an earlier state of a system by reversing its evolution, as seen in quantum mechanics, wave physics, and computer vision.
• Practical implementations range from heralded quantum state inversion using interferometric protocols to phase conjugation in wave propagation and backward sampling in algorithmic processes.
• Applications include quantum information processing with >95% fidelity, echo formation in water-wave and photonic experiments, and pre-capture frame recovery in computer vision studies.

Time rewinding denotes a family of procedures that reconstruct, emulate, or infer an earlier state of a dynamical process from a later one. In the literature, the term has several technically distinct meanings. In quantum information it can mean a heralded implementation of the inverse propagator $W^{-s}=e^{+iH_0 s\Delta T}$ for an arbitrary two-level system without prior knowledge of $H_0$ or of the repeatable interaction $V$ (Trillo et al., 2022). In wave physics it ranges from broadband phase conjugation, $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$, to temporal-modulation protocols that restore an evolved field to its exact original amplitude and phase (Mounaix et al., 2019, Kim et al., 19 Aug 2025). In computer vision and statistical simulation it can denote retrodictive reconstruction of pre-capture frames or historical trigger fields from present observations and auxiliary data (Chen et al., 2024, Luo et al., 2020). Across these usages, the common motif is reversal of an evolution law or of its observable consequences, but the operative meaning depends on the underlying formalism.

1. Conceptual forms of rewinding

In quantum mechanics, time reversal is tied to anti-unitarity. One formulation defines an anti-unitary operator $T$ by $T\rho T^\dagger:=K\rho K$, where $K$ is complex conjugation in the energy basis, and emphasizes that no purely unitary map reproduces $t\to -t$ for all observables; an anti-unitary step is essential (Lebedev et al., 2019). This is conceptually different from protocols that realize an inverse channel or inverse unitary by auxiliary control, postselection, or heralding.

For continuously monitored quantum systems, reversibility is conditioned on the measurement record itself. For a qubit with $H=(\hbar\Omega/2)\sigma_y$ and continuous $\sigma_z$ measurement, the readout is written as $H_0$0, and the backward movie is generated by reversing time order and negating the record, $H_0$1 (Dressel et al., 2016). The resulting backward evolution is physically possible, yet a statistical arrow of time remains because forward and backward hypotheses have different likelihoods.

In lossless linear wave propagation, time reversal is expressed by symmetry of the wave equation under $H_0$2, $H_0$3. In the frequency domain this gives $H_0$4, so practical time reversal becomes broadband phase conjugation of each spectral component (Mounaix et al., 2019). A related spacetime-transformation picture treats a sudden temporal boundary as the dual of a spatial mirror: the frequency changes sign, $H_0$5, while the wavevector is preserved (Bacot et al., 2015).

The term is also used algorithmically. In the Markov-chain and language-model settings, rewinding means returning to previously observed states and resuming stochastic evolution from them, rather than physically reversing microscopic dynamics (Azarmehr et al., 17 Feb 2026, Azarmehr et al., 24 Mar 2026). This suggests that “time rewinding” is best understood as a family resemblance term spanning exact inversion, echo production, backward sampling, and retrodictive reconstruction.

2. Universal rewinding of unknown qubit dynamics

A universal qubit rewinding protocol can be built from a two-path interferometric primitive. The target qubit is prepared together with a motion degree of freedom in the superposition $H_0$6. On path $H_0$7 the qubit undergoes free evolution $H_0$8 followed by the unknown but repeatable operation $H_0$9, while on $V$0 the order is reversed. After recombination on a balanced beam splitter, the heralded branch $V$1 applies the commutator $V$2, and the branch $V$3 applies the anticommutator $V$4. This primitive is the gate $V$5 and obeys

$V$6

For any $V$7 matrices $V$8, the identities $V$9, $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$0 for invertible $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$1, and $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$2 permit synthesis of the word $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$3, thereby rewinding the target by $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$4 (Trillo et al., 2022).

The protocol is universal in the sense specified for arbitrary two-level systems: it does not depend on the form of $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$5 or on the detailed interaction, provided that the same unknown $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$6 can be repeated. Its classical control layer is a random walk on a finite word-graph. If each application of $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$7 yields $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$8 with probability $E_{\rm rev}(\mathbf r,t)=E^*(\mathbf r,-t)$9, then in the generic case $T$0 one has

$T$1

and the failure probability decays exponentially fast in the number of trials. The protocol fails only in the fine-tuned commuting case $T$2, and the algebraic argument is specific to dimension two because it relies on $T$3 Cayley–Hamilton identities (Trillo et al., 2022).

A photonic realization used a quantum-SWITCH architecture to implement the required commutator structure and demonstrated reversal of discrete polarization evolution with an average state fidelity of over $T$4. For $T$5 free-evolution steps, the reported fidelities were

$T$6

with a grand mean of approximately $T$7 (Schiansky et al., 2022). The same work states that the protocol is optimal in running time: if the goal is to rewind by $T$8 discrete steps, at least $T$9 uses of the unknown free evolution are necessary.

The principal limitations are explicit. The target must be a two-level system, the unknown interaction must be repeatable across runs, and each use of the building block requires interferometric control of the motion degree of freedom together with one free-evolution interval $T\rho T^\dagger:=K\rho K$0 and one interaction step $T\rho T^\dagger:=K\rho K$1 (Trillo et al., 2022).

3. Measurement reversal, ancilla-assisted inversion, and optimal-control undo

Continuous quantum measurement supplies a distinct notion of rewinding. In the Kraus-operator description, the forward update over $T\rho T^\dagger:=K\rho K$2 is

$T\rho T^\dagger:=K\rho K$3

while the reversed trajectory is generated by

$T\rho T^\dagger:=K\rho K$4

with the record transformed as $T\rho T^\dagger:=K\rho K$5. Although every forward trajectory has a corresponding backward movie, the arrow of time appears statistically through the log-likelihood ratio

$T\rho T^\dagger:=K\rho K$6

When $T\rho T^\dagger:=K\rho K$7, the run is more likely to be forward than reversed (Dressel et al., 2016). The same framework generalizes to non-projective measurements by “Janus” sequences, where the backward Kraus operators satisfy $T\rho T^\dagger:=K\rho K$8.

A different route to reversing an unknown quantum state assumes knowledge of the Hamiltonian and uses an ancilla governed by the same Hamiltonian. With the SWAP operator $T\rho T^\dagger:=K\rho K$9 and partial-SWAP unitary

$K$0

the reduced channel after tracing out the ancilla is

$K$1

For small $K$2, this induces an effective Hamiltonian proportional to $K$3, and after $K$4 iterations the approximation error is bounded by $K$5 (Lebedev et al., 2019). The resource estimate given there scales as $K$6 for fidelity at least $K$7.

Experimental “undo” operations have also been realized by optimal control in a five-level system: the $K$8 Zeeman manifold of non-interacting $K$9Rb atoms on an atom chip. There the forward evolution under

$t\to -t$0

is approximately inverted by an optimally designed control pulse $t\to -t$1 satisfying $t\to -t$2. Using the dCRAB algorithm with a truncated Fourier-like basis of $t\to -t$3 complex modes, the experiment achieved on average an accuracy of around $t\to -t$4 across tens of test operations of duration $t\to -t$5, and an arbitrary-time undo protocol reached $t\to -t$6 overlap for reversal to an intermediate past state (Mastroserio et al., 2022). The Loschmidt echo,

$t\to -t$7

was used there as a thermodynamic measure of reversibility.

These formulations clarify a frequent confusion. Some quantum rewinding schemes invert unknown dynamics universally but only for qubits; some reverse stochastic state-update trajectories conditioned on a recorded measurement stream; others require a known Hamiltonian or optimal-control synthesis. Exact reversal, heralded reversal, and approximate undo are therefore separate regimes rather than interchangeable descriptions.

4. Wave-based rewinding: phase conjugation, temporal boundaries, echoes, and deterministic restoration

In optics, time-reversed waves are “pre-scattered” spatiotemporal fields that enter a complex medium as complicated inputs and arrive as prescribed targets. A system built from a $t\to -t$8-D spatial light modulator, multi-plane light conversion, and multimode fibre synthesized arbitrary vector spatiotemporal fields over an optical bandwidth $t\to -t$9, with spectral resolution $H=(\hbar\Omega/2)\sigma_y$0 per SLM pixel, $H=(\hbar\Omega/2)\sigma_y$1 Hermite–Gaussian modes in each of two polarizations, and approximately $H=(\hbar\Omega/2)\sigma_y$2 spatiotemporal degrees of freedom. Prescribed targets such as a $H=(\hbar\Omega/2)\sigma_y$3 spatiotemporal focus, delayed polarization-structured images, and volumetric “arrow of time” and Eiffel-tower patterns were reconstructed with more than $H=(\hbar\Omega/2)\sigma_y$4 correlation in both space and time (Mounaix et al., 2019).

An alternative wave-mechanical mechanism is the Instantaneous Time Mirror. Here the wave speed is made time dependent through a sudden, spatially uniform jump,

$H=(\hbar\Omega/2)\sigma_y$5

which converts the homogeneous wave equation into

$H=(\hbar\Omega/2)\sigma_y$6

The source term is interpreted as a distribution of “Cauchy sources” created everywhere at the temporal boundary. In water-wave experiments, a point impact launched an expanding packet, the bath was jolted after $H=(\hbar\Omega/2)\sigma_y$7 with acceleration reaching $H=(\hbar\Omega/2)\sigma_y$8 in about $H=(\hbar\Omega/2)\sigma_y$9, and a converging packet refocused at the source at $\sigma_z$0 (Bacot et al., 2015).

For nonrelativistic matter waves, a quantum time mirror was proposed as a near-instantaneous nonlinear kick

$\sigma_z$1

In the $\sigma_z$2-kick limit this imprints the phase

$\sigma_z$3

leaving the density unchanged but modifying the current according to

$\sigma_z$4

The resulting reversal is partial: in one dimension the norm-overlap reached about $\sigma_z$5 for $\sigma_z$6, $\sigma_z$7, and $\sigma_z$8, while in a two-dimensional ring geometry echoes as strong as $\sigma_z$9 were reported (Reck et al., 2018).

More recent work isolates a deterministic regime in time-varying media. In electromagnetic systems, carefully paired temporal layers with impedance matching or anti-matching and matched durations produce complete restoration of amplitude and phase; in Dirac systems the analogous condition is complete interband transition. The central claim is that, unlike time-reversal holography or quantum time mirrors, which produce wave echoes but only partial waveform recovery, the designed temporal modulation can achieve deterministic and complete reconstruction of the entire wave state (Kim et al., 19 Aug 2025). A symmetry classification sharpens this result: for any lossless, spatially homogeneous modulation sequence with identical initial and final media,

$H_0$00

with isotropic and chiral media being channel-preserving and Tellegen media channel-exchanging (Kim et al., 21 Jun 2026). This gives a direct scattering-matrix criterion for exact time rewinding.

5. Rewinding as an algorithmic primitive

In large-language-model inference, rewinding is modeled as interaction with a Markov chain in which the algorithm may resume generation from any previously observed state. The state space $H_0$01 consists of partial solutions, and a rewinding step chooses a prior state $H_0$02 and samples $H_0$03. The main theorem states that the optimal algorithm always generates a caterpillar tree: after removing the leaves of the explored state tree, what remains is a path. This yields the Caterpillar of Thoughts (CaT) algorithm (Azarmehr et al., 24 Mar 2026).

Theoretical characterization is accompanied by empirical comparisons. On $H_0$04 hard Game of 24 instances, Tree-of-Thoughts with beam $H_0$05 achieved $H_0$06 success at $H_0$07 average tokens, whereas CaT with best-of-$H_0$08 and $H_0$09 steps achieved $H_0$10 at $H_0$11 average tokens; with $H_0$12 steps, CaT achieved $H_0$13 at $H_0$14 (Azarmehr et al., 24 Mar 2026). On $H_0$15 held-out $H_0$16 crosswords, truncated ToT at $H_0$17 steps obtained word accuracy $H_0$18, letter accuracy $H_0$19, games solved $H_0$20, and $H_0$21 average tokens, while CaT at $H_0$22 steps obtained $H_0$23, $H_0$24, $H_0$25, and $H_0$26 respectively (Azarmehr et al., 24 Mar 2026).

Partially observable Markov chains admit a related but more abstract rewinding model. There the learner observes only $H_0$27 and may jump back according to a rewind function. Three strategy classes are distinguished: passive, adaptive rewinding, and non-adaptive rewinding. The central information-theoretic theorem states that if a pair of states can be distinguished by some adaptive rewinding strategy, then it can also be distinguished by a non-adaptive strategy. The efficiency difference appears only in query complexity, where a polynomial overhead for non-adaptive strategies is both achievable and necessary in general (Azarmehr et al., 17 Feb 2026).

Backward simulation of stochastic processes leads to Time-Reverse Monte Carlo. A naive “invert-and-simulate” procedure is biased unless Jacobian factors are included. The remedy is to introduce a backward proposal kernel $H_0$28 and incremental importance weights

$H_0$29

The full-path estimator

$H_0$30

is unbiased, and resampling in the SMC variant is more efficient for simulations with a larger number of time steps (Takayanagi et al., 2017).

Quantum complexity theory gives yet another formalization. Rewinding operators that invert quantum measurements define the class $H_0$31, and the main structural theorem is

$H_0$32

The same work shows that a single rewinding operator can already enable tasks believed intractable for quantum computation under standard assumptions, that rewindable Clifford circuits remain classically simulatable, and that rewindable IQP circuits can solve any problem in $H_0$33 (Hiromasa et al., 2022). In this algorithmic literature, rewinding is not a physical inversion of time but an added control primitive with measurable computational power.

6. Retrodiction, reconstruction, and reinterpretation

In computer vision, TimeRewind studies the problem of recovering the moments just before a single captured image. The inputs are a single RGB frame $H_0$34 at time $H_0$35 and an event stream

$H_0$36

over the interval $H_0$37. A frozen image-to-video diffusion backbone is augmented with an Event Motion Adaptor (EMA) that predicts residuals at each U-Net block. Training minimizes

$H_0$38

using only EMA parameters as trainable variables (Chen et al., 2024). On held-out sequences from BS-ERGB, the reported results were: SVD, $H_0$39, $H_0$40, $H_0$41; E2VID+, $H_0$42, $H_0$43, $H_0$44; EVDI, $H_0$45, $H_0$46, $H_0$47; REFID*, $H_0$48, $H_0$49, $H_0$50; and TimeRewind, $H_0$51, $H_0$52, $H_0$53 (Chen et al., 2024). The present rewind window is about $H_0$54.

In geohazard modeling, “rewinding to the past” denotes posterior simulation of historical trigger scenarios rather than inversion of dynamics. A Bayesian generalized additive model is fitted to the 2017 Jiuzhaigou earthquake-induced landslide inventory with

$H_0$55

and a PGA term enters linearly through $H_0$56. Backward simulation then removes the 2017 PGA effect and injects a historical PGA field for each earlier scenario, generating ensembles of susceptibility maps by posterior sampling (Luo et al., 2020). The Jiuzhaigou model yielded ten-fold cross-validation AUC values all above $H_0$57, with median approximately $H_0$58, and a posterior mean for $H_0$59 of approximately $H_0$60 (Luo et al., 2020).

General relativity offers a more geometric use of the term. The type-D Kasner vacuum metric

$H_0$61

can be read in two opposite time orientations. With $H_0$62 increasing from zero, it describes an anisotropic cosmology emerging from a Big Bang; with $H_0$63 decreasing toward zero, it describes the late interior of a Schwarzschild black hole approaching a future singularity (Robson et al., 2022). The same line element therefore supports both the “birth of a universe” and the “end of time” by reversal of time ordering, illustrating that in relativistic contexts rewinding may refer to reinterpretation of a solution rather than to an operational protocol.

Taken together, these strands show that time rewinding is not a single doctrine but a technically stratified concept. In some settings it is exact inverse dynamics; in others it is heralded inversion, echo formation, optimal-control undo, backward sampling, state-space backtracking, or probabilistic reconstruction. A plausible implication is that the most useful distinctions are not between disciplines but between guarantees: exact versus approximate restoration, deterministic versus heralded success, microscopic reversibility versus statistical arrow, and physical inversion versus inferential retrodiction.