Shadow Lindblad Tomography

• Shadow Lindblad tomography is a set of techniques that reconstructs Lindblad operators from compressed dynamical data instead of conventional informationally complete measurements.
• It leverages both a dynamical shadow approach from a single observable’s time trace and a randomized classical-shadow protocol to estimate local Lindblad parameters under locality assumptions.
• These methods enable efficient extraction of coherent and dissipative dynamics in quantum systems, dramatically reducing experimental configurations and scaling costs compared to standard process tomography.

Shadow Lindblad tomography denotes a family of tomographic schemes for open quantum systems in which a Lindblad generator, or observables determined by it, are reconstructed from compressed dynamical data rather than from a conventional catalog of separately programmed informationally complete measurements. In the cited literature, the expression appears in two closely related senses. One is a “dynamical shadow” picture in which the full state is encoded in the time trace of a single observable evolving under a known Lindblad semigroup. The other is a randomized classical-shadow protocol that estimates local Lindblad parameters from short-time Pauli-transfer data under locality assumptions. Both viewpoints treat open-system dynamics not merely as a nuisance but as a structured source of tomographic information (Rall et al., 17 Jan 2025, Birke et al., 16 Feb 2026).

1. Formal setting and reconstructed objects

The common starting point is a Markovian master equation

$$\dot{\rho} \;=\; \mathcal{L}(\rho) \;=\; -\frac{i}{\hbar}[H,\rho] + \sum_k L_k\rho L_k^\dagger - \frac{1}{2}\{L_k^\dagger L_k,\rho\},$$

with $H$ the Hamiltonian and $\{L_k\}$ jump operators. In Lindblad tomography, the object of inference is typically the time-independent generator itself: either $H$ together with a Lindblad or Kossakowski matrix in a fixed operator basis, or an equivalent jump-operator representation (Samach et al., 2021).

For $n$-qubit systems, one convenient parameterization expands the generator in the Pauli basis $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$: $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$ The coherent coefficients $a_P$ and dissipation-matrix entries $D_{P,Q}$ are the Lindblad parameters in the experimental shadow protocol of 2026. In the fully general case there are $4^n-1$ coherent parameters and $H$0 real parameters in $H$1, for a total of $H$2, which is the source of the standard $H$3 scaling barrier (Birke et al., 16 Feb 2026).

A short-time formulation makes the inverse problem linear. If $H$4 denotes an expectation value for input observable $H$5 and output observable $H$6, then

$H$7

Collecting configurations gives

$H$8

and, after choosing $H$9 such that $\{L_k\}$0 on the parameter sector of interest, one obtains transformed expectation values $\{L_k\}$1 whose slopes at $\{L_k\}$2 equal the Lindblad parameters (Birke et al., 16 Feb 2026).

This linearization underlies both extensible Lindblad tomography and its shadow variants. The distinction is not the model but the way the required expectation values are acquired.

2. Dynamical-shadow tomography from a single observable

A distinct line of work shows that full quantum tomography can, in principle, be performed from the homogeneous time evolution of a single expectation value. For a Hilbert space $\{L_k\}$3, a known homogeneous evolution $\{L_k\}$4, and one fixed observable $\{L_k\}$5, the measured signal is

$\{L_k\}$6

or, in discrete time,

$\{L_k\}$7

The central result is that, for generic dynamics, a single nontrivial binary measurement record is informationally complete for state tomography (Rall et al., 17 Jan 2025).

In discrete time, if $\{L_k\}$8, then for every nontrivial $\{L_k\}$9 and for almost all quantum channels $H$0, the map

$H$1

is injective on $H$2. The same generic injectivity holds when the channel is restricted to a semigroup element $H$3. The continuous-time corollary states that for almost every Lindbladian $H$4, and for $H$5 distinct times $H$6, the map

$H$7

is injective for every fixed nontrivial $H$8. The exceptional set has measure zero in the natural $H$9-dimensional parameter space of generators (Rall et al., 17 Jan 2025).

The same work proves a quantum analogue of Takens’ embedding theorem. If $n$0 is a prior-constrained closed set of states and $n$1, with $n$2 the Minkowski dimension, then for almost all channels $n$3 the delay map

$n$4

is injective on $n$5 and has a $n$6-Hölder continuous inverse on its image for any $n$7. This incorporates prior information directly into the tomography criterion. For example, the rank-$n$8 density matrices in dimension $n$9 have dimension $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$0, so prior rank information reduces the required number of delay coordinates (Rall et al., 17 Jan 2025).

A central structural point is that nontrivial noise is not merely tolerated but required. The paper shows that unitary evolution, even with added simply depolarizing noise, is insufficient beyond the qubit case. By contrast, generic Lindbladians possess distinct decay rates and oscillation frequencies, and the expectation signal takes the form

$\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$1

with coefficients $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$2 that can span the operator space when the spectrum is nondegenerate. This is the precise sense in which the time record functions as a low-dimensional “dynamical shadow” of the state (Rall et al., 17 Jan 2025).

3. Classical-shadow reconstruction of local Lindbladians

The 2026 experimental formulation of shadow Lindblad tomography applies classical-shadow ideas directly to Lindblad-parameter estimation. Rather than reconstructing a state from a single time series, it estimates many low-weight Pauli-transfer-matrix elements from one randomized data stream and then converts their short-time slopes into Lindblad parameters (Birke et al., 16 Feb 2026).

For $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$3 qubits, the relevant channel statistics are

$\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$4

where $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$5 and $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$6 denotes idling evolution under the Lindbladian. Each shadow experiment samples a random product Pauli eigenstate as input, lets the device idle for time $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$7, and measures in a random product Pauli basis. If $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$8, $\mathcal{P}_n=\{I,X,Y,Z\}^{\otimes n}$9, with supports $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$0 and $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$1, the single-shot estimator is

$$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$2

when the random preparation and measurement axes agree with all non-identity factors of $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$3 and $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$4, and $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$5 otherwise. Averaging over $$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$6 randomizations gives

$$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$7

The estimator is unbiased,

$$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$8

and satisfies

$$H = \sum_{P\in \mathcal{P}_n} a_P\, P,\qquad \mathcal{L}= -i\sum_{P\in\mathcal{P}_n} a_P [P,\rho] + \sum_{P, Q \in \mathcal{P}_n} D_{P,Q}\left(P\rho Q - \frac{1}{2}\{QP,\rho\}\right).$$9

Because each single-shot contribution is bounded in magnitude by $a_P$0, Hoeffding concentration applies directly (Birke et al., 16 Feb 2026).

The practical gain comes from locality. In a $a_P$1-local Lindbladian model, only low-weight $a_P$2 contribute to the parameter sector of interest, so the required estimator variance is independent of the total system size for fixed $a_P$3. The paper states that, under fixed $a_P$4, the number of physically relevant parameters grows polynomially with $a_P$5, while the number of randomizations needed to estimate each low-weight Pauli-transfer element to fixed accuracy grows only logarithmically with the number of parameters. This is the basis of the claimed $a_P$6 sample complexity behavior under locality assumptions (Birke et al., 16 Feb 2026).

This randomized protocol sits naturally within a broader shadow-process framework based on the Choi isomorphism. A channel $a_P$7 can be represented by its Choi state $a_P$8, and shadow process tomography estimates quantities of the form

$a_P$9

from randomized input and output unitaries. That formalism supplies channel-level shadow estimators, sample-complexity bounds, and composition rules for channels, and therefore provides a general process-tomography backdrop for Lindblad-specific shadow schemes (Kunjummen et al., 2021).

4. Experimental realizations and benchmarks

The experimental precursor to shadow Lindblad tomography is full Lindblad tomography on superconducting hardware. On a small transmon processor, Lindblad tomography was used to reconstruct a time-independent Lindblad generator from time-domain measurements while explicitly calibrating state preparation and measurement errors. The protocol modeled the dissipator through a positive semidefinite Lindblad matrix $D_{P,Q}$0 in a Pauli-like basis,

$D_{P,Q}$1

and extracted both Hamiltonian terms and effective jump operators from maximum-likelihood fits across many time points (Samach et al., 2021).

In that experiment, single- and two-qubit reconstructions identified conventional dephasing and amplitude-damping channels, quantified crosstalk, and resolved an always-on ZZ term

$D_{P,Q}$2

with $D_{P,Q}$3. The same study used the Breuer–Laine–Piilo trace-distance criterion to distinguish genuine non-Markovianity from effective subsystem non-Markovianity induced by tracing out another qubit. Its central scaling conclusion was that full Lindblad tomography, like process tomography and gate-set tomography, remains exponentially costly beyond small subsystems (Samach et al., 2021).

The 2026 benchmark converted this framework into a shadow protocol on a five-qubit superconducting transmon processor. It first implemented extensible Lindblad tomography as a baseline and then compared it to shadow Lindblad tomography on one- and three-qubit subsystems. On these subsystems, the shadow protocol reproduced extensible Lindblad tomography within uncertainties while using exponentially fewer configurations. For the three-qubit case, extensible Lindblad tomography reconstructed all $D_{P,Q}$4 parameters, and the observed hierarchy showed dominant 1-local terms, smaller 2-local terms, and 3-local coherent and incoherent parameters less than one standard deviation from zero, supporting a 2-local model (Birke et al., 16 Feb 2026).

That locality verification enabled the five-qubit deployment. Under a 2-local model, shadow Lindblad tomography recovered all single-qubit dissipation and two-qubit coupling parameters in 9 hours of acquisition time, compared to an estimated 58 hours for extensible Lindblad tomography under the same restricted model. The paper also states that unrestricted five-qubit extensible tomography, with $D_{P,Q}$5 parameters, would be infeasible in comparable conditions and was estimated at roughly 162 days. A further practical result is statistical: the shadow estimator is compatible with conventional Gaussian error propagation, so the analysis did not require median-of-means estimators (Birke et al., 16 Feb 2026).

A common misconception is that “shadow” methods remove all exponential dependence automatically. The benchmarked protocol does not make that claim. Its efficiency gain is explicitly tied to physically motivated locality assumptions and to the fact that only low-weight Pauli-transfer data are needed (Birke et al., 16 Feb 2026).

5. Gate-level error tomography and cyclic error amplification

A separate but related direction concerns tomography of Lindbladian gate errors rather than idle-system generators. Robust Lindbladian Tomography (RLT) models each implemented gate as

$D_{P,Q}$6

with $D_{P,Q}$7 the Lindbladian error, and analyzes repeated sequences of cyclic gates through error amplification circuits. The objective is to infer the gate-wise errors $D_{P,Q}$8 while accounting for non-commutativity between ideal and error generators and between different gates in the sequence (Sugiyama, 16 Mar 2025).

The main technical device is a first-order perturbative expansion that is exact to all orders in the ideal generator $D_{P,Q}$9 and first order in the small error $4^n-1$0. For diagonalizable $4^n-1$1, the paper defines linear maps $4^n-1$2, $4^n-1$3, $4^n-1$4, and $4^n-1$5 through spectral projectors and coefficients

$4^n-1$6

and proves decomposition and composition formulas for $4^n-1$7 and products of noisy gates. For a cyclic unit repeated $4^n-1$8 times, the effective error splits into an amplified part

$4^n-1$9

and a non-amplified part

$H$00

so that angle-like errors scale linearly with repetition number $H$01, whereas axis-like errors do not (Sugiyama, 16 Mar 2025).

RLT uses this structure to transform the reconstruction problem into a convex semidefinite program. The fitted variables are the Lindbladian error matrices $H$02, the forward model is linear in these variables after perturbative reduction, trace preservation is imposed by

$H$03

and complete positivity is encoded through the linear matrix inequality

$H$04

The paper emphasizes that this reduces the numerical optimization from a nonlinear fit to an SDP, although the overall cost still grows exponentially with the number of qubits, as in other tomographic methods (Sugiyama, 16 Mar 2025).

This is not, by itself, a classical-shadow protocol. However, the paper explicitly notes that its linearized inverse problem could be combined with shadow-style front ends: linear estimators of channel functionals obtained from randomized measurements could be inserted into the same linear or SDP post-processing. A plausible implication is that gate-level shadow Lindblad tomography would inherit its basic geometry from RLT and its sample-efficiency gains from classical shadows (Sugiyama, 16 Mar 2025).

6. Non-Markovian extensions, statistical limits, and constraints

Shadow Lindblad tomography is usually formulated for time-independent Markovian generators, but related work has extended Lindblad-like learning to non-Markovian time-local master equations. In Lindblad-like quantum tomography, the generator is written in time-local form

$H$05

with the key difference that the rates $H$06 are allowed to become negative. For single-qubit dephasing,

$H$07

and Ramsey data at multiple times are fitted jointly through a likelihood over the full time series rather than through separate snapshots. The work studies both frequentist maximum-likelihood and Bayesian sequential Monte Carlo inference, and shows that optimal measurement times depend strongly on the correlation time and on the degree of non-Markovianity. In the non-Markovian regime, Bayesian adaptive scheduling outperforms the corresponding frequentist strategy in the numerical comparisons reported there (Varona et al., 2024).

The same paper quantifies non-Markovianity through the CP-divisibility measure

$H$08

and a trace-distance measure $H$09 that, for dephasing, is determined by intervals where $H$10. This provides a direct reminder that Lindblad-parameter learning and shadow-based compression do not, by themselves, settle the Markovianity question; the dynamical model must still be validated against the underlying time correlations (Varona et al., 2024).

A second constraint comes from quantum metrology. For estimation of weak Lindblad decay rates under arbitrary fast and precise control, the optimal strategy is to rapidly projectively measure and re-initialize the quantum state. In the vanishing-signal limit, the optimal quantum Fisher information for $H$11 is

$H$12

and it scales linearly with total interrogation time $H$13. The corresponding precision obeys standard-quantum-limit scaling rather than Heisenberg scaling. This result applies even in the full fast-control model with ancillae, and the paper proves that a measure-and-reset strategy saturates the ultimate bound (Gardner et al., 6 Jan 2025).

For shadow Lindblad tomography, this metrological result does not prescribe a unique reconstruction algorithm, but it does bound what any algorithm can extract from repeated short-time data. A plausible implication is that randomized classical shadows can improve front-end data reuse and reduce configuration count, while the total information available about weak Lindblad rates remains constrained by the same $H$14 law (Gardner et al., 6 Jan 2025).

Two limitations therefore recur across the literature. First, exact knowledge or accurate modeling of the generator is often assumed; model misspecification, drift, or leakage can bias inference. Second, efficiency gains depend on structure: locality, low-weight observables, prior constraints, cyclicity, or low-dimensional parameterizations. Outside those regimes, shadow Lindblad tomography reverts toward the same fundamental scaling difficulties that affect general process tomography (Samach et al., 2021, Kunjummen et al., 2021).