Process Tensor Method: Multi-time Quantum Dynamics

• Process Tensor Method is defined as a high-order tensor mapping sequences of quantum interventions to final states, effectively capturing non-Markovian memory.
• It utilizes tensor-network techniques such as PT-MPOs and temporal MPSs to efficiently simulate multi-time dynamics in open quantum systems.
• Applications include two-dimensional spectroscopy, quantum error correction under correlated noise, and tensor-based process control in semiconductor manufacturing.

Searching arXiv for relevant papers on process tensors and their recent applications. The process tensor method denotes, in open quantum systems, a multi-time representation of reduced dynamics in which a system $S$ coupled to an environment $E$ is described by a high-order object $\Upsilon_{k:0}$ that maps a sequence of interventions to a final state or measurement record. In this usage, the process tensor is a multilinear map, a Choi operator, and a quantum comb, and it is designed to retain non-Markovian memory that is inaccessible to single-step CPTP maps or Markovian master equations. Recent work has realized this object numerically through tensor-network constructions such as PT-MPOs and temporal MPSs, with applications to two-dimensional electronic spectroscopy, quantum error correction under correlated noise, and full counting statistics in interacting transport (Wit et al., 2024, Kobayashi et al., 2024, Yadalam et al., 30 Mar 2026). The same label also appears in a distinct semiconductor-manufacturing literature, where it refers instead to tensor-based process modeling, control, and monitoring for image-valued disturbances (Li et al., 2024).

1. Multi-time dynamical object

In the quantum setting, the process tensor is introduced for an $n$-step experiment in which an experimenter applies completely-positive maps $\mathcal A_k$ to a system at discrete times while the system and environment interact unitarily between interventions. The central claim is that the multi-time influence of the environment on the system is captured by a single object,

$$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$

which acts on the sequence of interventions and returns the final state on the output space. In de Wit et al., the same object is described as a rank-$2k$ tensor that takes preparations, measurements, and control operations at times $t_0,t_1,\dots,t_{k-1}$ and returns the final system state at time $t_k$ (Wit et al., 2024). In the transport setting of Yadalam and Mitchison, it is also identified with the influence functional or influence matrix seen by an interface subsystem embedded in a many-body environment (Yadalam et al., 30 Mar 2026).

A key structural point is that the process tensor is a generalization of a completely positive trace-preserving map from one time step to many. For $k=1$, a single-step CPTP map is recovered. For $E$0, $E$1 encodes joint dependence on multiple interventions and thus retains history dependence. This is the precise sense in which it captures non-Markovianity: the environment is not summarized by an instantaneous channel but by a higher-order object containing correlations across time (Wit et al., 2024).

2. Choi-state, quantum-comb, and contraction formalisms

The standard formal representation is Choi–Jamiołkowski. If $E$2 is a sequence of CP maps with Choi states $E$3, the final output state is written as

$E$4

where $E$5 is the link product, i.e. a partial trace and index contraction. In vectorized form, the same relation becomes

$E$6

This establishes the process tensor as the Choi operator of an $E$7-step map from a sequence of instruments to a final state, which is why it is also described as a quantum comb (Kobayashi et al., 2024).

The object can also be written by feeding halves of maximally entangled pairs through the joint system-environment dynamics and tracing out the environment. In de Wit et al., one formal expression is

$E$8

with $E$9. A more compact open-system expression writes

$\Upsilon_{k:0}$0

Once $\Upsilon_{k:0}$1 is known, multi-time response functions follow by inserting the desired superoperators into its time slots and contracting the resulting tensor network (Wit et al., 2024).

A common misconception is that a process tensor is merely a sequence of two-time maps. In the strictly Markovian limit, it does factorize as

$\Upsilon_{k:0}$2

so each slot is independent. Outside that limit, failure of such factorization is exactly the signature of non-Markovian memory in the formalism (Wit et al., 2024).

3. Tensor-network realizations and computational structure

Because the full object scales exponentially with the number of times, practical use depends on tensor-network compression. In the spectroscopy work of de Wit et al., time is discretized with step $\Upsilon_{k:0}$3, the Gaussian bath is expressed in Feynman–Vernon path-integral form, and the time-evolving matrix product operator (TEMPO) algorithm represents the discrete influence functional as a growing tensor network composed of bath tensors $\Upsilon_{k:0}$4 that connect system states $\Upsilon_{k:0}$5 steps apart. Half-step system propagators are written as $\Upsilon_{k:0}$6, and a bath memory cutoff $\Upsilon_{k:0}$7 neglects longer-range tensors when $\Upsilon_{k:0}$8. Singular-value truncation with tolerance $\Upsilon_{k:0}$9 then compresses the network into an MPO, yielding a PT-MPO of length $n$0. The cost to build the PT grows roughly as $n$1, where $n$2 is the typical MPO bond dimension, and once the PT is built, each additional time argument costs only an extra linear sweep through the MPO (Wit et al., 2024).

In the decoder setting, both the process tensor and tester are likewise approximated by one-dimensional tensor networks. The process tensor is written as an MPO of length $n$3 with bond dimension $n$4, and tester or recovery objects are written as MPOs or MPSs of comparable bond dimension. The contraction $n$5 then reduces to an MPO/MPS contraction of length $n$6, executable in $n$7 time by sequential contraction and singular-value truncation. This approximation is explicitly limited by the requirement that the process entanglement entropy remain low; exact tomography of $n$8 grows exponentially in $n$9, and contraction ordering and bond-dimension choice are critical to the accuracy–cost tradeoff (Kobayashi et al., 2024).

In the many-body transport formulation, the process tensor is recast as a temporal MPS. A distinctive technical feature is normalization-preserving truncation. The exact object satisfies a no-intervention normalization condition,

$\mathcal A_k$0

A naive SVD truncation generally violates this condition, so Yadalam and Mitchison perform an exact SVD, contract future and past environments into tall matrices, use QR decompositions to identify the singular-value subspace required by normalization, and truncate only the complementary subspace. The stated error bound is given by the sum of the discarded singular values squared, while normalization is preserved exactly (Yadalam et al., 30 Mar 2026).

4. Spectroscopic response and transport statistics

The spectroscopy application in de Wit et al. targets two-dimensional electronic spectroscopy of non-Markovian open quantum systems. In the rotating-wave, impulsive, and phase-matched limits, the relevant observables are the third-order response functions

$\mathcal A_k$1

In the PT language, each dipole interaction is inserted as a left or right superoperator, $\mathcal A_k$2 or $\mathcal A_k$3, at the appropriate time slot of the PT-MPO. The final 2D signal is assembled as

$\mathcal A_k$4

so the spectrum is obtained by two Fourier transforms over the first and third time delays with waiting time $\mathcal A_k$5 set to zero. The implementation reported by de Wit et al. is in the open-source OQuPy package and yields numerically exact 2D spectra with full non-Markovian and temperature-dependent bath effects (Wit et al., 2024).

The transport application in Yadalam and Mitchison uses the process tensor of the two sites straddling a cut in a one-dimensional many-body lattice system with $\mathcal A_k$6 symmetry. The moment-generating function of transferred charge is obtained from a repeated insertion of a counting-field-tilted local superoperator,

$\mathcal A_k$7

Because $\mathcal A_k$8 is independent of $\mathcal A_k$9, it is constructed once and reused to scan $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$0. Benchmarks on the Heisenberg spin-$$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$1 XXZ brickwork circuit at infinite temperature recover ballistic, superdiffusive, and diffusive regimes. The reported scaling forms are $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$2 and $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$3 in the ballistic regime $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$4, $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$5 and $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$6 at the isotropic point $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$7, and $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$8 and $$\Upsilon_{n:0}\in \mathscr B\!\left(\bigotimes_{k=0}^{n-1}(H_k^{\rm in}\otimes H_k^{\rm out})\right),$$9 in the diffusive regime $2k$0. The same work reports a self-similar scaling form of the generating function outside the ballistic regime and confirms the breakdown of Kardar-Parisi-Zhang universality in higher-order transport cumulants at the isotropic point (Yadalam et al., 30 Mar 2026).

5. Quantum error correction with spatiotemporally correlated noise

In quantum error correction, the process tensor formalism is used to represent noise that is not iid but instead contains non-trivial spatiotemporal correlations, including cross-talk, non-Markovianity, and mixtures of the two. The decoder framework examined in the strategic-code setting constructs probabilities of classical syndrome records directly from a process tensor. If $2k$1 is the measurement record and $2k$2 is the tester Choi state, then for a hypothesized error process $2k$3,

$2k$4

Logical recovery is then formulated as maximum-likelihood decoding. If $2k$5 denotes a recovery Choi operator and $2k$6 the combined tester-recovery object, the decoder chooses

$2k$7

This makes the decoder explicitly dependent on the full multi-time noise model rather than on per-time or per-qubit marginals (Kobayashi et al., 2024).

The factorized iid Pauli model appears as a limiting special case. There, the process tensor factorizes,

$2k$8

the required tensor-network rank is $2k$9, and standard products of local error probabilities are recovered. By contrast, the full process tensor allows arbitrary temporal and spatial correlations generated by coupling the system to a bath of arbitrary size. The numerical study reported by Kalfus et al. emphasizes both the detrimental effects of correlated noise and the possibility of decoder designs that account for such effects, while also noting the computational limitations of exact process-tensor tomography and the dependence of low-bond-dimension approximations on weakly non-Markovian structure (Kobayashi et al., 2024).

6. Distinct usage in semiconductor process control

A separate 2024 literature uses “Process-Tensor Method” in a non-quantum sense. In Li et al., the term refers to tensor-based process control and monitoring for semiconductor manufacturing with unstable disturbances. The basic run-to-run model is

$t_0,t_1,\dots,t_{k-1}$0

where $t_0,t_1,\dots,t_{k-1}$1 is the control-recipe adjustment, $t_0,t_1,\dots,t_{k-1}$2 is an unobserved first-type disturbance, $t_0,t_1,\dots,t_{k-1}$3 is a second-type uncompensable disturbance, and $t_0,t_1,\dots,t_{k-1}$4 is the image-based overlay error tensor. The coefficient tensor $t_0,t_1,\dots,t_{k-1}$5 is given a Tucker structure,

$t_0,t_1,\dots,t_{k-1}$6

with low multilinear ranks $t_0,t_1,\dots,t_{k-1}$7. Parameter estimation is carried out either by least squares or by a group-lasso-plus-ridge problem (GLRP), solved through alternating updates and block-coordinate descent (Li et al., 2024).

Control is then implemented by an EWMA law in tensor space. With

$t_0,t_1,\dots,t_{k-1}$8

the core-space prediction error is

$t_0,t_1,\dots,t_{k-1}$9

and the next control action is

$t_k$0

Li et al. state a stability theorem in which the EWMA law is asymptotically stable if and only if all eigenvalues of $t_k$1 satisfy $t_k$2, and a corollary reduces this to $t_k$3 under a specific bias model for $t_k$4. Residual monitoring is performed through a tensor residual $t_k$5, with $t_k$6, $t_k$7, and EWMA charts defined in projected tensor coordinates. The reported results include a simulation in which the uncontrolled case has $t_k$8 and EWMA control gives $t_k$9 for $k=1$0, as well as lithography results in which the tensor-EWMA controller achieves roughly $k=1$1–$k=1$2 reduction in average overlay error relative to existing image-based feedback control, especially under autocorrelated or non-stationary disturbances (Li et al., 2024).

This suggests that the phrase “Process Tensor Method” is not semantically uniform across arXiv literatures. In quantum theory it denotes a multi-time dynamical map or its Choi-state/tensor-network realization; in semiconductor process engineering it denotes tensor-valued process modeling, control, and monitoring. The shared element is the use of high-order tensor structure to capture dependencies that are not well represented by lower-order or factorized models, but the underlying objects, observables, and computational goals are different (Li et al., 2024).