Temporal State Tomography (TST)

• Temporal State Tomography (TST) is a framework that treats time as an inherent state coordinate, enabling the recovery of temporal density matrices and multi-time operators.
• It employs advanced techniques such as frequency-offset homodyne detection, polarization-resolved interference, and dispersive time-bin POVMs, achieving high reconstruction fidelities (up to 0.99).
• TST’s diverse implementations across quantum optics and classical imaging provide unified insights into temporal coherence, phase structure, and dynamic system evolution.

Temporal State Tomography (TST) denotes a family of reconstruction procedures in which temporal structure is part of the state description rather than a mere acquisition label. In the literature, the term is not standardized. It can refer to the reconstruction of a temporal density matrix or temporal mode function of a narrowband single photon, tomography of time-bin or temporal-mode photonic qudits, reconstruction of a multi-time quantum process as a single temporal state $\Upsilon$, time-resolved density-matrix tomography of classical optical beams, and, in broader non-quantum usage, temporal characterization of dynamical systems such as evolving tomographic volumes or active-matter clusters (Yang et al., 2018, Sedziak-Kacprowicz et al., 2020, Jia, 4 May 2026, Plöschner et al., 2022, Luzzatto et al., 12 Nov 2025). The common denominator is that time enters the reconstructed object itself.

1. Terminological scope and reconstructed objects

The surveyed literature uses “Temporal State Tomography” for several technically distinct tasks. In quantum optics, TST often means recovering phase-sensitive temporal information that is inaccessible to direct counting, such as the complex temporal mode function $\varphi(\tau)$, the temporal density matrix $\rho_{TM}$, or a finite-dimensional density operator in a time-bin or temporal-mode basis (Yang et al., 2018, Chen et al., 2014, Sedziak-Kacprowicz et al., 2020). In recent quantum-information formalisms, TST instead denotes tomography of a multi-time process represented by a temporal state $\Upsilon$, with temporal quasiprobability distributions as the experimentally accessible data model (Jia, 4 May 2026, Jia et al., 8 Jan 2026). Outside that setting, the same terminology or closely related language appears in dynamic CT, spatiotemporal scattering tomography, and classical high-dimensional optical field reconstruction (Nikitin et al., 2018, Ronen et al., 2020, Plöschner et al., 2022).

Domain	Reconstructed object	Operational content
Multi-time quantum processes	Temporal state $\Upsilon$, TQD, temporal KD/MH data	Unified state-and-channel tomography
Single-photon temporal-spectral optics	$\rho_{TM}$, $\varphi(\tau)$	Homodyne/heterodyne temporal coherence reconstruction
Biphoton temporal tomography	$A(\tau)e^{i\phi(\tau)}$	Two-photon interference with polarization projections
Time-bin and temporal-mode photonics	Finite-dimensional density matrix	Dispersive POVM or QPG-based projective measurements
Classical spatiotemporal optics	$\hat\rho(\lambda,t)$	Time- and wavelength-resolved density matrices
Broader dynamical tomography	$f(x,y,z,t)$, $\varphi(\tau)$0, $\varphi(\tau)$1	Dynamic reconstruction or temporal characterization

A recurring consequence of this terminological breadth is that “state” does not always mean a positive semidefinite single-time density operator. In some works it means a temporal-spectral density matrix, in others a multi-time operator, and in still others a compact descriptor of temporal organization.

2. Multi-time quantum process formulations

A recent line of work formulates TST as tomography of a quantum process across multiple times, reconstructed as a single temporal object rather than as separately estimated states and channels (Jia, 4 May 2026). For an $\varphi(\tau)$2-time process

$\varphi(\tau)$3

the reconstructed object is a temporal state $\varphi(\tau)$4. This object encodes both the initial state and the intervening dynamics. Unlike ordinary density operators, temporal states are generally not positive semidefinite, which makes direct transplantation of standard tomography nontrivial (Jia, 4 May 2026).

The operational solution is based on temporal quasiprobability distributions (TQDs). In the informationally complete setting, a TQD uniquely determines the temporal state. A central construction expresses the TQD as a quasiprobability over trajectories through a temporal phase space,

$\varphi(\tau)$5

The paper then shows that one need not directly implement the generally nonphysical phase-space operations $\varphi(\tau)$6. Instead, one can measure a fixed quantum instrument sequentially and reconstruct the TQD by classical post-processing,

$\varphi(\tau)$7

For informationally complete TQDs, the temporal state can then be reconstructed by a dual-frame map such as

$\varphi(\tau)$8

or by a temporal Bloch-type expansion (Jia, 4 May 2026).

The sample-complexity analysis in the same framework yields

$\varphi(\tau)$9

and, in the natural scaling $\rho_{TM}$0,

$\rho_{TM}$1

so the asymptotic statistical cost matches ordinary informationally complete tomography up to the dimension of the temporal operator space (Jia, 4 May 2026).

A closely related 2026 formulation extends the Kirkwood–Dirac quasiprobability distribution to arbitrary multi-time quantum processes and uses left, right, and doubled temporal KD quasiprobabilities, together with their real temporal Margenau–Hill counterparts, as the operational basis for temporal or spatiotemporal Bloch tomography (Jia et al., 8 Jan 2026). In that framework, the temporal Born rule takes forms such as

$\rho_{TM}$2

and the doubled KD object is the most information-complete. The same paper explicitly relates KD temporal states, MH temporal states, pseudo-density operators, doubled density operators, process tensors, quantum combs, quantum strategies, and superdensity operators within one operational scheme (Jia et al., 8 Jan 2026). This establishes TST not merely as a measurement protocol but as a unifying representation theory for multi-time quantum data.

3. Correlation-based and learning-based variants

Before the explicit temporal-state formalisms, process reconstruction from temporal data had already appeared in a correlation-based form. A 2012 method reconstructs quantum dynamics from temporal correlations between observables, rather than from preparing many known input states and performing state tomography on the outputs (Ber et al., 2012). The central measured object is the two-time temporal covariance matrix

$\rho_{TM}$3

To access these correlations without strongly disturbing the system, the method uses weak measurements with pointers. For discrete finite-dimensional systems, if $\rho_{TM}$4 is invertible, the Heisenberg-picture affine map is reconstructed by

$\rho_{TM}$5

with the affine shift $\rho_{TM}$6 obtained from one-point averages. The same work shows complexity $\rho_{TM}$7 for a $\rho_{TM}$8-dimensional discrete system, while for Gaussian channels the method becomes state-independent and scales as $\rho_{TM}$9 for $\Upsilon$0 particles (Ber et al., 2012). Its distinctive feature is applicability to mixed and thermal states without controlled state preparation.

A different direction, aimed at memory-bearing devices, is “learning temporal quantum tomography” (Tran et al., 2021). There the object is not a memoryless channel but a temporal map whose output at time $\Upsilon$1 depends on input history. The method uses a quantum reservoir with recurrent update

$\Upsilon$2

followed by measurement of observables $\Upsilon$3, producing features

$\Upsilon$4

A linear readout $\Upsilon$5 is then trained to reconstruct vectorized output density matrices. Reconstruction quality is quantified by the fidelity

$\Upsilon$6

and by the root mean square of fidelities,

$\Upsilon$7

The paper reports RMSF above $\Upsilon$8 in the shown examples for moving-average and delayed-depolarizing tasks, and introduces a quantum short-term memory capacity,

$\Upsilon$9

as a measure of temporal processing ability (Tran et al., 2021). This is an approximate, supervised-learning realization of TST for temporally correlated quantum devices rather than an exact informationally complete reconstruction scheme.

4. Temporal-spectral tomography of single photons and biphotons

In narrowband photonic experiments, TST is primarily a phase-sensitive temporal-spectral reconstruction problem. For heralded subnatural-linewidth single photons generated by spontaneous four-wave mixing in a cold $\Upsilon$0 atomic cloud, ordinary photon counting reveals only $\Upsilon$1, whereas TST reconstructs the full temporal mode function $\Upsilon$2 and temporal density matrix $\Upsilon$3 (Yang et al., 2018). In a discrete time-bin basis,

$\Upsilon$4

The key measured quantity is the reduced autocorrelation matrix,

$\Upsilon$5

where $\Upsilon$6. Homodyne detection at $\Upsilon$7 gives the real part directly, while heterodyne detection with multiple local-oscillator offsets reconstructs both real and imaginary parts (Yang et al., 2018). The experiment used eight LO frequency offsets,

$\Upsilon$8

and reported temporal purity

$\Upsilon$9

The imaginary part was found to be very small, and the phase approximately constant over the $\rho_{TM}$0 ns coherence window (Yang et al., 2018).

For narrowband biphotons, the target is the full complex temporal waveform

$\rho_{TM}$1

rather than only the coincidence envelope $\rho_{TM}$2 (Chen et al., 2014). The method uses six polarization-resolved two-photon interference measurements after a beam splitter and polarization analyzers to recover the phase difference function

$\rho_{TM}$3

for $\rho_{TM}$4, with $\rho_{TM}$5. The phase is then reconstructed recursively. To bridge separated amplitude islands while retaining high temporal resolution, the experiment employed a two-step recursion with

$\rho_{TM}$6

The reported reconstructions recovered $\rho_{TM}$7-phase jumps at amplitude nodes, frequency-induced phase steps in nondegenerate cases, and a Fourier-transform-limited waveform in a longer-coherence regime (Chen et al., 2014).

These photonic realizations clarify one of the oldest meanings of TST: it is the recovery of temporal coherence and phase structure that are invisible to direct arrival-time statistics.

5. Time-bin, temporal-mode, and compressive measurement architectures

Three experimentally distinct families illustrate how TST depends on measurement design: dispersive time-resolved POVMs for time-bin qudits, calibrated quantum pulse gates for temporal-mode states, and randomized compressive tomography in the time-frequency domain (Sedziak-Kacprowicz et al., 2020, Ansari et al., 2017, Gil-Lopez et al., 2021).

For time-bin photonic qudits, one approach lets the photon propagate through a dispersive fiber and then performs time-resolved single-photon detection (Sedziak-Kacprowicz et al., 2020). Dispersion mixes otherwise separated bins in a known, state-independent way, so the measurement is described by a time-dependent POVM,

$\rho_{TM}$8

with detector jitter incorporated by convolution with a Gaussian response. The reconstruction uses least squares and maximum likelihood estimation, with physicality enforced via

$\rho_{TM}$9

The paper numerically tested $\varphi(\tau)$0 qubit states, $\varphi(\tau)$1 qutrits, and $\varphi(\tau)$2 entangled states. Representative average fidelities include $\varphi(\tau)$3 for qubits with $\varphi(\tau)$4 m and $\varphi(\tau)$5 using least squares, and $\varphi(\tau)$6 for qubits with $\varphi(\tau)$7 m and $\varphi(\tau)$8 ps using least squares, while longer fiber improved performance under large jitter (Sedziak-Kacprowicz et al., 2020).

For temporal-mode tomography, the quantum pulse gate (QPG) is treated as a programmable mode-selective detector whose measurement operators must first be characterized (Ansari et al., 2017). In a finite-dimensional Hermite–Gaussian temporal-mode basis, the effective operator for pump setting $\varphi(\tau)$9 is

$A(\tau)e^{i\phi(\tau)}$0

Measurement tomography is performed with a tomographically complete set of mutually unbiased bases. The paper reports $A(\tau)e^{i\phi(\tau)}$1 measurements for $A(\tau)e^{i\phi(\tau)}$2 and $A(\tau)e^{i\phi(\tau)}$3 for $A(\tau)e^{i\phi(\tau)}$4. After calibration, state reconstruction of unknown temporal-mode states is carried out with the experimentally determined $A(\tau)e^{i\phi(\tau)}$5. The reported calibrated state-tomography fidelities are $A(\tau)e^{i\phi(\tau)}$6 for 5D unfiltered, $A(\tau)e^{i\phi(\tau)}$7 for 5D filtered, and $A(\tau)e^{i\phi(\tau)}$8 for 7D, substantially improving over reconstructions that assume ideal measurement operators (Ansari et al., 2017).

A more explicitly compressive variant realizes universal compressive tomography in the time-frequency domain using a QPG that implements arbitrary projections by pump-pulse shaping (Gil-Lopez et al., 2021). The reconstructed object is a low-rank temporal pulsed-mode or frequency-bin state in a finite $A(\tau)e^{i\phi(\tau)}$9-dimensional subspace. The protocol uses randomly rotated orthonormal bases sampled from the Haar measure, maximum-likelihood preprocessing of relative frequencies, and informational completeness certification based on the convex-set indicator

$\hat\rho(\lambda,t)$0

If $\hat\rho(\lambda,t)$1, the measured bases are informationally complete. The demonstration used $\hat\rho(\lambda,t)$2 Hermite–Gaussian modes and $\hat\rho(\lambda,t)$3 frequency bins, and reported that the number of measured bases required for completeness, $\hat\rho(\lambda,t)$4, increases roughly linearly with the rank $\hat\rho(\lambda,t)$5 (Gil-Lopez et al., 2021). The same paper emphasizes that the method avoids ad hoc assumptions about the exact rank, sparsity, or state family of the unknown signal.

6. Broader classical and non-quantum extensions

The concept of temporal-state reconstruction also appears in classical optics. A 2022 method reconstructs a density matrix $\hat\rho(\lambda,t)$6 of an optical beam as a function of wavelength and time by using a spatial light modulator to display projective holograms, a single-mode fibre as a single-pixel projection channel, and either a high-speed photodiode or a spectrometer as the resolved detector (Plöschner et al., 2022). For each temporal or spectral bin, the measured projective intensities yield a generalized Stokes vector, from which the density matrix is reconstructed. The eigenvalues give the weights of mutually incoherent spatial components, while the eigenvectors give the corresponding spatial fields, including amplitude and phase. In the experimental $\hat\rho(\lambda,t)$7 basis, the full tomographic set required

$\hat\rho(\lambda,t)$8

analyser masks. Temporal resolution was set by a $\hat\rho(\lambda,t)$9 photodiode and spectral resolution by a $f(x,y,z,t)$0 optical spectrum analyser in the example (Plöschner et al., 2022).

In X-ray and scattering tomography, the same broad idea appears as dynamic or spatiotemporal tomography rather than quantum-state estimation. One 4D reconstruction method models the object as

$f(x,y,z,t)$1

and solves

$f(x,y,z,t)$2

using a Chambolle–Pock primal-dual algorithm accelerated on GPU hardware (Nikitin et al., 2018). Its purpose is to relax the traditional assumption that the sample remains static during one tomographic rotation. A related cloud-imaging study derives CT of a time-varying volumetric translucent object using a small number of moving cameras, focusing on passive scattering tomography of dynamic clouds, discussing required angular and temporal sampling rates, and solving the inverse problem by gradient-based optimization (Ronen et al., 2020).

An even broader, explicitly non-quantum use of “temporal tomography” occurs in active matter (Luzzatto et al., 12 Nov 2025). There, temporal cluster tomography is defined through the time intervals $f(x,y,z,t)$3 between successive events in which a chosen particle pair belongs to the same cluster. The resulting temporal gap-size statistics $f(x,y,z,t)$4 are summarized by the burstiness parameter

$f(x,y,z,t)$5

The paper states directly that this is not a full-state reconstruction method in the quantum tomography sense. Rather, it is a statistical temporal state characterization of non-equilibrium clustered matter (Luzzatto et al., 12 Nov 2025).

7. Methodological themes, misconceptions, and limitations

A common misconception is that TST always reconstructs a positive semidefinite density operator at each time. In the multi-time process literature, the reconstructed temporal state is generally not positive semidefinite, and this non-positivity is one reason quasiprobabilistic or Bloch-type reconstructions are introduced (Jia, 4 May 2026). Another source of confusion is terminological: temporal cluster tomography in active matter is explicitly not the quantum-informational meaning of state tomography, even though it reconstructs a temporal descriptor of dynamical state (Luzzatto et al., 12 Nov 2025).

Across the surveyed quantum implementations, informational completeness is achieved in markedly different ways. Some schemes rely on multiple LO frequency settings to separate real and imaginary parts of temporal coherence (Yang et al., 2018); some use six polarization-resolved interference measurements and recursive phase reconstruction (Chen et al., 2014); some design time-dependent POVMs via dispersive propagation (Sedziak-Kacprowicz et al., 2020); some require detector tomography and calibration of the measurement operators themselves (Ansari et al., 2017); and some use random Haar bases with convex certification of completeness (Gil-Lopez et al., 2021). In the multi-time process setting, informational completeness is framed in terms of IC-POVMs, dual frames, and fixed quantum instruments followed by classical post-processing (Jia, 4 May 2026).

The principal limitations are likewise domain-specific. For time-bin tomography, detector jitter strongly affects the effective POVM, and longer fiber is not always universally better because excessive dispersion changes the measurement geometry (Sedziak-Kacprowicz et al., 2020). For QPG-based methods, selectivity, mode matching, phasematching asymmetries, side lobes, pump-shaping imperfections, and acquisition time constrain performance, even though calibration can raise state-tomography fidelity to about $f(x,y,z,t)$6 (Ansari et al., 2017). Universal compressive tomography is most efficient for low-rank or near-coherent states and still requires finite-dimensional truncation and numerical informational-completeness certification (Gil-Lopez et al., 2021). In correlation-based process tomography, weak measurements introduce both random and systematic error, and the discrete-variable method requires a nonsingular initial state for full reconstruction (Ber et al., 2012).

Taken together, these results indicate that TST is less a single protocol than a recurring reconstruction pattern: time is promoted to a native coordinate of the state description, and experimental design is then organized around how to make that temporal object observable. In some settings the object is a complex temporal waveform, in others a finite-dimensional density matrix, and in the most general recent formalisms it is a multi-time operator that unifies state and channel tomography within one scheme (Chen et al., 2014, Yang et al., 2018, Jia, 4 May 2026).