Process-Tensor Formulation

• Process-Tensor Formulation is a framework that captures multi-time dynamics in open quantum and high-dimensional classical systems by mapping sequences of interventions to final states.
• It employs tensor-network methods, such as matrix-product operators, to compress high-dimensional data and reduce computational complexity from exponential to polynomial scaling.
• The formulation generalizes classical stochastic kernels and quantum dynamical maps, effectively addressing non-Markovian memory effects and enabling advanced simulation and control.

The process-tensor formulation is a rigorous framework for capturing the multi-time dynamics of open quantum systems and high-dimensional processes, generalizing both classical stochastic kernels and quantum dynamical maps to accommodate non-Markovian memory, external interventions, and tensorial data structures. By encoding the causal response of a system under a sequence of interventions or controls, the process tensor provides a mathematical object—typically a multi-linear superoperator or a tensor network—that unifies disparate simulation, control, and modeling approaches in quantum physics, process engineering, and complex networks. This formalism has become foundational for the analysis, simulation, and optimal control of non-Markovian dynamics, and also for high-dimensional regression and monitoring in classical process systems.

1. Origin and Mathematical Definition

The process-tensor formalism extends the operator-algebraic theory of quantum stochastic processes, originally developed by Accardi, Frigerio, and Lewis (AFL), by embedding non-commutative multi-time correlation kernels into a Schrödinger-picture object that maps a sequence of completely positive (CP) interventions to final system states (Nurdin et al., 2021). For a quantum system coupled to an environment, with discrete time steps $k=0,1,\ldots,N$, each system intervention $U_k$ (a CP map) is interleaved with propagations of the total system-bath state. The reduced system state at the final time is written as

$$\rho_S(N) = \Upsilon_{N:0}\left[U_N,\ldots,U_1\right](\rho_S(0)),$$

where $\Upsilon_{N:0}$ (“process tensor”) is a unique multi-linear map or superoperator, constructed by projecting the full chain of open-system dynamics onto the relevant bath subspaces (Ortega-Taberner et al., 2024, Wit et al., 2024).

With the Choi–Jamiołkowski isomorphism, $\Upsilon_{N:0}$ admits a representation as a positive operator (the “Choi state” or “quantum comb”) on $\mathcal{H}_S^{\otimes 2N}$, and for classical Markov processes collapses to a chain of transition kernels (Nurdin et al., 2021).

2. Tensor-Network Construction and Compression

The process tensor, as defined directly, is exponentially costly with respect to the number of time steps due to exponentially growing memory in non-Markovian environments. To address this, the process tensor is efficiently encoded as a matrix-product operator (MPO) in Liouville space (Ortega-Taberner et al., 2024, Wit et al., 2024): $$\Upsilon^{(\mu_N,\nu_N),\ldots,(\mu_1,\nu_1)} = \sum_{\chi_0\cdots\chi_N} O[1]^{\mu_1,\nu_1}_{\chi_1,\chi_0} O[2]^{\mu_2,\nu_2}_{\chi_2,\chi_1}\cdots O[N]^{\mu_N,\nu_N}_{\chi_N,\chi_{N-1}},$$ where $O[k]$ are local MPO tensors with physical legs corresponding to system indices and auxiliary “bond” legs $\chi_{k-1},\chi_k$ encoding environmental memory.

Several simulation methods map naturally into this PT-MPO structure:

• TEMPO/path-integral-MPO (OQuPy): constructs MPOs by sequential singular-value decompositions and truncations on the Feynman–Vernon influence functional, adaptively compressing to memory length and bath correlation time.
• HEOM: represents the process tensor with bond dimension equal to the number of auxiliary modes.
• Stochastic Liouville/von Neumann: averages over $M$ noise trajectories, yielding a PT-MPO with bond $\chi=M$.
• Augmented system/pseudomodes: couples the system to $A$ discrete auxiliary oscillators, with bond $\chi=\mathrm{dim}(A)$.

This compression enables polynomial rather than exponential scaling with respect to simulation time, allowing exact non-Markovian simulation and optimal control gradient computation (Ortega-Taberner et al., 2024, Wit et al., 2024).

3. Applications: Quantum Control, Spectroscopy, and High-dimensional Process Modeling

Quantum Optimal Control

In open quantum systems, the PT-MPO provides a unified abstraction to design and evaluate optimal control protocols under non-Markovian noise. Back-propagation through the PT-MPO yields gradients with respect to interventional maps (controls), using forward (“state”) and backward (“costate”) recursions within the MPO structure (Ortega-Taberner et al., 2024). This supports end-to-end optimization of control fields, generalizes GRAPE/adjoint algorithms, and allows direct comparison of solver efficiency via MPO bond dimension.

Multi-time Correlations and 2D Spectroscopy

The process-tensor formalism provides the exact multi-time response functions required for nonlinear spectroscopy. Arbitrary multi-time correlations are computed by inserting operator superoperators at desired time-slices within the PT-MPO and contracting the network (Wit et al., 2024). In 2D electronic spectroscopy, this enables calculation of rephasing and nonrephasing pathways without Born–Markov approximations, with polynomial numerical scaling controlled by MPO bond dimension and SVD truncation threshold.

Tensorial Process Models in Engineering and Data Science

Recent works generalize the process-tensor concept to high-dimensional process control and machine learning:

• Tensor-based process control: Models such as $$\mathcal{Y}_t = \mathcal{B}\times_1(u_t+d_t) + \mathcal{E}_t$$ relate control vectors and disturbances to tensor-valued outputs, with the process tensor $\mathcal{B}$ parameterized by Tucker decompositions. Control and monitoring are performed directly in the reduced core space, while residuals monitor non-compensable disturbances (Li et al., 2024).
• Tensor-on-tensor regression neural networks: Deep architectures preserve multilinear structure by stacking Tucker layers and core contraction (Einstein product), with backpropagation exploiting the process-tensor mapping between high-dimensional input and outputs. This structure is crucial for tractable modeling of high-dimensional sensor, image, and point-cloud data (Wang et al., 6 Oct 2025).

4. Links to Operator-algebraic Stochastic Processes and Classical Limits

The process-tensor arises from the AFL theory of quantum stochastic processes by specializing their multi-time correlation kernels (in the Heisenberg picture) to sequential CP maps and tracing out environments, possibly with ancilla augmentation to reproduce all instrument statistics (Nurdin et al., 2021). Classically, commutative subalgebras and quantum non-demolition measurements yield consistency relations that recover standard multi-time probability distributions and Markov transition chains.

Non-Markovianity is precisely encoded by correlations (typically quantum entanglement) in the Choi state across time-slots. The process tensor is quantum Markovian if and only if the Choi state is tensor-factorized. Otherwise, its internal structure reveals the full extent of quantum memory and correlations induced by the environment.

5. Complexity, Scalability, and Algorithmic Implementation

The computational complexity of constructing and utilizing the process tensor is dominated by the bond dimension $\chi$ of the PT-MPO. For a system of Hilbert space dimension $S$ and maximal bond $\chi_d$, each MPO tensor is of size $(S^2\times S^2)\times (\chi_d\times\chi_d)$. Forward and backward passes (for state and costate propagation) scale as $O(S^4\chi_d^2)$ per time step, with total scaling $O(N S^4 \chi_d^2)$ (Ortega-Taberner et al., 2024, Wit et al., 2024). Compression via SVD on MPO bonds is governed by the environmental memory time and the chosen truncation error $\epsilon_{\rm rel}$. For Markovian evolutions, $\chi=1$: the process tensor reduces to a chain of local propagators.

Method	Bond dimension $\chi$	Cost per step	Notes
TEMPO (full)	$S^{2M}$	Exponential in memory length	Without MPO compression
PT-MPO (TEMPO)	Adaptive	$O(S^4\chi_d^2)$	Compression keeps $\chi_d$ small
HEOM	$N_{\mathrm{aux}}$	$O(N_{\mathrm{aux}}S^4)$	Exploits sparsity
Transfer-tensor	$T_C$	$O(N T_C S^4)$	Finite memory-cutoff approach

Algorithmic construction involves sequential contraction of bath-influence tensors, SVD truncation, and stepwise assembly of the MPO. For optimal control and deep regression applications, end-to-end backpropagation leverages the process-tensor structure for efficient gradient computation (Ortega-Taberner et al., 2024, Wang et al., 6 Oct 2025).

6. Extensions beyond Quantum Systems: Hetero-functional Graphs and Multilayer Networks

In modeling large-scale engineering systems or multicomponent infrastructures, the process-tensor concept encompasses higher-order adjacency, capability, and incidence tensors. Hetero-functional graph theory (HFGT) employs a 4th-order adjacency tensor $\mathcal{A}_p$ to encode feasible sequences of process-resource pairs and 3rd/4th-order incidence tensors $\mathcal{M}^{\pm}$ for operand flow (Farid et al., 2021). These constructions relate directly to multilayer network representations, with tensorial slicing and projection revealing mappings that standard matrix or graph-based formalisms cannot capture.

The process-tensor perspective thus systematically unifies open-system quantum dynamics, high-dimensional process modeling in engineering, and advanced network theory, leveraging tensor structures to expose causal, memory, and control structure across disciplines.