Dynamical Quantum Tomography

• Dynamical quantum tomography is the study of reconstructing quantum states or processes by leveraging time-evolution to expand the effective measurement span.
• Protocols use known dynamics and limited measurement setups to achieve informational completeness, online tracking, and robust state estimation via iterative updates.
• Recent approaches extend tomography to processes, Liouvillian generators, and many-body systems, addressing challenges like noise resilience, derivative estimation, and scalability.

Dynamical quantum tomography denotes a family of quantum-characterization protocols in which temporal structure is part of the inverse problem. In the most common formulation, an unknown state is subjected to a known evolution and probed by a fixed or restricted measurement setup at multiple times, so that dynamics enlarge the effective measurement span and can replace a large static tomographically complete measurement family. In other formulations, the unknown state itself changes during data acquisition and is tracked online, or the reconstructed object is not a state but a process, a Liouvillian, a Hamiltonian, a reduced subsystem state, or a family of dynamical correlation functions (Kech, 2016, Rambach et al., 2022, Aguiar et al., 14 Apr 2025).

1. Formal definitions and informational completeness

A standard finite-dimensional formulation uses a POVM $P=(Q_1,\dots,Q_m)$ together with a known Heisenberg-picture evolution $\mathcal T$. The associated dynamical measurement scheme is

$$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$

with measurement data

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$

Within this framework, a measurement scheme is $\mathcal R$-complete when it is injective on a restricted state set $\mathcal R$, and informationally complete when it is injective on the full state space. For a feasible unitary $U\in U(n)$, almost all POVMs with $n+1$ outcomes yield informationally complete schemes $\mathcal T_U^n(P)$; with prior information $\mathcal R$ represented by a semi-algebraic difference set $\mathcal T$0, the condition $\mathcal T$1 guarantees stable $\mathcal T$2-completeness for almost all $\mathcal T$3-outcome POVMs, while feasible invertible CPTP dynamics remove the unitary lower bound $\mathcal T$4 and permit two-outcome schemes (Kech, 2016).

A complementary control-theoretic formulation defines the observable subspace

$\mathcal T$5

for a known Heisenberg evolution $\mathcal T$6 and an available observable family $\mathcal T$7. Full dynamical quantum state tomography is feasible exactly when

$\mathcal T$8

For Markovian dynamics, this reduces to finite Krylov-type tests: $\mathcal T$9 in continuous time, or

$$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$0

in discrete time, with deterministic rank tests built from observability matrices $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$1 and $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$2 (Peruzzo et al., 29 Sep 2025). This same idea appears in open-system stroboscopic tomography for phase-damping channels, where the evolution

$$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$3

is decomposed as

$$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$4

so that repeated measurements at times $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$5 yield a linear system for the projections $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$6; reconstruction requires $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$7 and

$$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$8

with $$\mathcal T^l(P)=\Big((Q_1,\dots,Q_m),(\mathcal T(Q_1),\dots,\mathcal T(Q_m)),\dots,(\mathcal T^{l-1}(Q_1),\dots,\mathcal T^{l-1}(Q_m))\Big),$$9 (Czerwinski et al., 2015).

A more specialized channel-based construction uses a time-dependent average channel

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$0

built from Weyl–Heisenberg operators $$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$1. Measuring one fixed projector $$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$2 at $$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$3 time instants gives a linear system whose unknowns are

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$4

If the Gram matrix

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$5

is nonsingular, the projectors $$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$6 are informationally complete, and the corresponding normalized family is an IC-POVM (Cao et al., 2024).

2. Online state tracking and evolving-state tomography

A distinct branch of dynamical tomography addresses the case in which the unknown state changes while data are being acquired. In this setting, batch estimators such as maximum-likelihood estimation and least-squares tomography are poorly matched to the experiment because they require a tomographically complete data set before returning an estimate. Matrix-exponentiated gradient tomography instead updates the estimate after each measurement record: $$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$7 with single-measurement loss

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$8

and, in the photonic implementation, a multi-outcome version

$$h_{\mathcal T^l(P)}(x)=\big(\operatorname{tr}(Q_i(\mathcal T^\dagger)^j(x))\big)_{\substack{i=1,\dots,m\ j=0,\dots,l-1}}.$$9

Because the update acts through $\mathcal R$0 and matrix exponentiation, positivity and unit trace are preserved at every step, so the estimate remains a valid density matrix without ad hoc projection (Rambach et al., 2022).

The experimental realization was carried out on a photonic qutrit encoded in transverse spatial modes. Measurements were chosen sequentially from informationally complete families based on mutually unbiased bases or generalized Pauli operator measurements, and the evolving target state was taken as

$\mathcal R$1

with stationary, structured, and random-Hermitian trajectories. With a constant learning rate $\mathcal R$2, the protocol reached infidelity below $\mathcal R$3 in about 4 iterations in high-count qutrit experiments and in roughly 13–16 iterations for low-count MUB tomography, while mean infidelities remained around $\mathcal R$4, corresponding to fidelities around $\mathcal R$5. Under added background noise up to approximately $\mathcal R$6 kHz, performance stayed comparable to the no-added-noise case within uncertainty bars, and the authors interpret this as corresponding to $\mathcal R$7 (Rambach et al., 2022).

A related pure-state line assumes a known time-independent Hamiltonian and projective intensity measurements, treating reconstruction as phase retrieval under Schrödinger evolution and aiming to decrease the number of distinct projectors by exploiting the known unitary dynamics (Czerwiński, 2015). This suggests that “dynamical” can refer either to online tracking of a drifting state or to using known evolution as an informational resource even when the target state itself is static.

3. Process, channel, and Liouvillian tomography

Dynamical tomography frequently targets the evolution law itself rather than an instantaneous state. One early route is weak-measurement-based process tomography, where a process

$\mathcal R$8

is expanded as

$\mathcal R$9

and each coefficient $\mathcal R$0 is linked directly to a weak-value-like quantity $\mathcal R$1. In the proposed scheme, every process parameter is determined from only five experimental values—four pointer correlations $\mathcal R$2 and one post-selection probability $\mathcal R$3—and complete tomography requires only $\mathcal R$4 setups, with product input states sufficient even for multiparticle processes (Zhang et al., 2013).

Another route keeps the object of reconstruction at the channel level but uses tomography to expose dynamical structure invisible to scalar coherence times. In superconducting-qubit dynamical decoupling experiments, single-qubit QPT in the $\mathcal R$5-matrix representation

$\mathcal R$6

showed that dynamical decoupling suppresses dephasing but does not suppress spontaneous emission, and that pulse imperfections generate coherent residual rotations under $\mathcal R$7-4. Robust sequences such as $\mathcal R$8-8, $\mathcal R$9-16), KDD, and UR20 remove the oscillatory behavior caused by control errors, showing that process tomography can distinguish dephasing suppression, persistent relaxation, and coherent pulse-induced artifacts within a single reconstructed dynamical map (Souza, 2020).

More recent work moves from maps to generators. Room-temperature qutrit process tomography in a $U\in U(n)$0 vapor reconstructs finite-time process matrices $U\in U(n)$1 in the Bloch–Fano basis and then infers an effective total relaxation superoperator $U\in U(n)$2, separating residual Zeeman Hamiltonian terms, dephasing, and isotropic relaxation in a realistic noisy ensemble (Sun et al., 27 Aug 2025). Lindblad-like quantum tomography then generalizes the semigroup paradigm to time-local non-Markovian maps by maximizing a multi-snapshot likelihood over time-local master equations with possibly negative decay rates,

$U\in U(n)$3

and shows explicitly, for single-qubit dephasing, why multiple temporal snapshots are required once the dynamics is not time homogeneous (Varona et al., 2024). Quantum Liouvillian Tomography pushes the same program to multi-qubit open dynamics by combining gradient-based QPT with regression over derivatives of Pauli-string probabilities,

$U\in U(n)$4

to reconstruct a time-local Liouvillian

$U\in U(n)$5

and to detect non-Markovianity through negative canonical rates in idle two-qubit superconducting dynamics (Aguiar et al., 14 Apr 2025).

4. Many-body, subsystem, and correlation-function tomography

In many-body settings, dynamical tomography is often feasible only because the target states occupy a structured submanifold of Hilbert space. Matrix product state tomography exploits the fact that out-of-equilibrium states of one-dimensional systems with finite-range interactions remain efficiently approximable by MPS for any fixed evolution time. The protocol reconstructs local reduced density matrices on contiguous blocks of size $U\in U(n)$6, requiring at most $U\in U(n)$7 local Pauli bases per $U\in U(n)$8-site block for qubits, and then fits a global MPS with a certified fidelity lower bound

$U\in U(n)$9

In a trapped-ion simulator, this enabled reconstruction of dynamical states of up to 14 spins, with $n+1$0 for the initial product state in the 8-spin experiment and $n+1$1 up to $n+1$2 ms for triplet-block tomography, while the eventual collapse of the certificate tracked the physical spread of correlations beyond the chosen block size (Lanyon et al., 2016).

A more targeted subsystem variant appears in lattice gauge theory, where quench dynamics in the Schwinger model were characterized not only through Loschmidt echoes and non-equal-time correlators but also through time-resolved entanglement tomography of reduced density matrices. The reconstructed subsystem state

$n+1$3

was analyzed through the second Rényi entropy

$n+1$4

and through a fitted entanglement Hamiltonian ansatz $n+1$5, making it possible to extract time-dependent entanglement spectra and entanglement Hamiltonians during a dynamical quantum phase transition (Mueller et al., 2022).

The same logic extends from states to families of observables. Fermionic-Adapted Shadow Tomography reformulates dynamical commutators and anti-commutators into expectation values compatible with shadow methods, so that many correlators can be estimated simultaneously with at most two-copy measurements and uncontrolled Hamiltonian simulation. For example, the commutator

$n+1$6

is reduced to ordinary expectations on three effective evolved states, while retarded Green’s functions

$n+1$7

are treated through probabilistically prepared branches $n+1$8. The resulting sample complexities improve from $n+1$9 to $\mathcal T_U^n(P)$0 or $\mathcal T_U^n(P)$1 in the commutator case, and from $\mathcal T_U^n(P)$2 to $\mathcal T_U^n(P)$3 in the anti-commutator regime $\mathcal T_U^n(P)$4 (Ko et al., 5 Aug 2025).

5. Spectroscopic and Hamiltonian-learning formulations

In ultrafast molecular spectroscopy, dynamical tomography is formulated as the reconstruction of a time-dependent molecular density operator from measured angularly resolved observables. A maximum-entropy approach represents the reconstructed state as

$\mathcal T_U^n(P)$5

or, for the NH$\mathcal T_U^n(P)$6 case study,

$\mathcal T_U^n(P)$7

The measured anisotropy coefficients

$\mathcal T_U^n(P)$8

link experimental photoelectron observables to molecular angular distribution moments, enabling time-resolved reconstruction of the lab-frame density matrix $\mathcal T_U^n(P)$9, visualization of charge migration, and extraction of the electronic subsystem entropy $\mathcal R$0 (Makhija et al., 2024).

A different spectroscopic variant targets the generator of the dynamics rather than the state: Hamiltonian learning in engineered quantum magnets from spatially resolved dynamical response. There the basic data are local spectral functions

$\mathcal R$1

measured across several impurity configurations and related to tunneling spectroscopy through

$\mathcal R$2

A supervised network then infers couplings $\mathcal R$3 in an effective spin Hamiltonian with Heisenberg exchange, anisotropy, and Dzyaloshinskii–Moriya interaction. In the noiseless case, the reported fidelities are $\mathcal R$4 for $\mathcal R$5, $\mathcal R$6, and $\mathcal R$7, and $\mathcal R$8 for $\mathcal R$9; with noise strength $\mathcal T$00, the multi-impurity protocol improves the corresponding fidelities to $\mathcal T$01, $\mathcal T$02, $\mathcal T$03, and $\mathcal T$04, respectively (Karjalainen et al., 21 Oct 2025).

6. Conceptual boundaries, limitations, and open problems

The term “dynamical quantum tomography” is not uniform across the literature. Some work uses “dynamical” in the modern sense of exploiting physical time evolution to gain informational completeness or to track a changing state, while other work casts the reconstruction algorithm itself as a dynamical system. In the latter category, pure-state tomography can be reformulated through a physical imposition operator whose fixed points are the states compatible with the measured distributions; in that framework, multiple compatible reconstructions appear as bifurcations, but the underlying tomography problem remains static rather than time-resolved (Goyeneche et al., 2014). A related distinction concerns the inferred object: state-tracking methods infer $\mathcal T$05, process and Liouvillian methods infer $\mathcal T$06 or $\mathcal T$07, and correlation-function or Hamiltonian-learning methods infer selected observables or model parameters rather than a full density matrix (Rambach et al., 2022, Aguiar et al., 14 Apr 2025, Karjalainen et al., 21 Oct 2025).

The main technical limitations are also formulation-dependent. Online MEG tomography demonstrates robust empirical tracking but does not derive new nonstationary tracking bounds, and its responsiveness depends on the learning-rate tradeoff between noise averaging and adaptation (Rambach et al., 2022). Time-local generator methods such as Lindblad-like tomography and QLT assume that the reduced dynamics admit a time-local master equation; they are therefore broader than semigroup tomography but narrower than fully general process-tensor reconstructions, and they are especially sensitive to derivative estimation and discretization choices (Varona et al., 2024, Aguiar et al., 14 Apr 2025). Structured many-body protocols gain efficiency only while entanglement and correlation length remain controlled; in MPS tomography, the block size $\mathcal T$08 must eventually grow with time, so the method is efficient in system size at fixed time but not generically efficient in evolution time itself (Lanyon et al., 2016).

Taken together, these results suggest three persistent design principles. First, temporal diversity can substitute for measurement diversity, but only when the known dynamics genuinely expands the observable span. Second, structural priors—low entanglement, restricted rank, operator sparsity, time-locality, or Hamiltonian parametrization—are not incidental conveniences but the main reason scalable dynamical tomography is presently possible. Third, open-system dynamics can enlarge tomographic reach in ways that unitary dynamics cannot: for time-homogeneous Markovian unitary evolution, a single nontrivial observable is insufficient when $\mathcal T$09, whereas generic open dynamics can make even single-observable tomography feasible (Peruzzo et al., 29 Sep 2025). This suggests that future progress will likely come less from a single universal protocol than from a continued stratification of tomography problems by the dynamical structure they exploit (Kech, 2016, Peruzzo et al., 29 Sep 2025).