Time-Evolving Matrix Product Operator (TEMPO)

• Time-Evolving Matrix Product Operator (TEMPO) is a tensor-network approach that simulates open quantum system dynamics by compressing influence functionals into MPS/MPO representations.
• It leverages path-integral mapping and SVD-based compression to capture non-Markovian memory effects in both bosonic and fermionic baths.
• Recent iGTEMPO advances exploit time-translational invariance and infinite boundary conditions to achieve scalable, numerically exact quantum impurity simulations.

The Time-Evolving Matrix Product Operator (TEMPO) method is a tensor-network approach for simulating the dynamics of open quantum systems interacting with large reservoirs. TEMPO exploits the structure of the Feynman–Vernon path integral, integrating out environmental degrees of freedom to obtain a non-Markovian influence functional, which is efficiently represented as a matrix product state/operator (MPS/MPO) in the time domain. Fermionic extensions, notably Grassmann TEMPO (GTEMPO), employ Grassmann algebra to encode the anticommuting nature inherent in fermionic systems. Recent developments, such as infinite Grassmann TEMPO (iGTEMPO), leverage underlying time-translational invariance and infinite boundary conditions to yield highly efficient algorithms for equilibrium and non-equilibrium quantum impurity problems.

1. Formalism: Path-Integral Mapping and Influence Functional Construction

The TEMPO methodology starts from the reduced dynamics of a quantum impurity coupled to a non-interacting bosonic or fermionic bath, where the environmental degrees of freedom can be traced out exactly due to Gaussian statistics. This yields an influence functional—commonly denoted $\mathcal{F}$—that encapsulates all non-Markovian memory effects through two-point kernel integrals. For fermionic systems, the Grassmann path integral formulation is used: trajectories $a_p(\tau)$, $\bar{a}_p(\tau)$ for impurity flavor $p$ are integrated over an appropriate Grassmann measure (Guo et al., 2024).

The discretized path integral on a time grid $\tau = j\delta\tau$, $j=1\dots N$ is mapped into a quadratic functional over the set of Grassmann modes: $$I_p[\bar{a}_p,a_p] \simeq \exp\left( - \sum_{j,k=1}^N a_{p,j} \Delta_{j-k} \bar{a}_{p,k} \right)$$ where the hybridization kernel $\Delta_{j-k}$ is typically fitted via Prony or Padé decomposition into a sum of exponentials. This structure enables direct mapping to a Grassmann-valued matrix product operator (GMPO) or matrix product state (GMPS) with controllable bond dimension (Chen et al., 2023, Chen et al., 2023).

For bosonic TEMPO, the same principles apply but with scalar fields, leading to similar MPS representations (Strathearn et al., 2017).

2. Matrix Product State/Operator Representation and Compression

The core innovation of TEMPO lies in compressing the high-dimensional augmented density tensor (ADT) resulting from time discretization into an MPS/MPO. Each time slice is associated with physical indices corresponding to forward/backward branches of the Keldysh contour (real time) or occupation variables (imaginary time). The influence functional's exponential structure allows factorization into partial two-site tensors of small bond dimension, which are multiplied sequentially with successive bond truncations via singular value decomposition (SVD).

For Grassmann-valued problems, the MPS represents functions of anticommuting variables and respects fermion parity. The GMPS construction embeds auxiliary Grassmann variables between adjacent sites and enforces local parity constraints, enabling efficient local contraction and compression (Xu et al., 2024). MPO exponentiation—critical for long-range bath memories—is realized by repeated squaring, which, complemented by Prony decomposition of the kernel, limits the number of necessary MPO multiplications independent of the evolution time (Guo et al., 2024).

3. Time-Translational Invariance, Infinite Boundaries, and the iGTEMPO Algorithm

At zero temperature or in steady-state regimes, both the influence functional and the impurity propagator exhibit time-translational invariance. The key benefit is the elimination of boundary effects and growing modes, so that the ADT can be represented as an infinite GMPS with a fixed, repeating unit cell (Guo et al., 2024, Guo et al., 2024). This causes the required bond dimension $\chi$ for convergence to saturate, removing scaling with the number of time slices $N$ that otherwise plagues finite-temperature implementations.

The infinite GTEMPO workflow comprises:

1. Discretization of $\delta\tau$ and MPO construction for the local propagator and kernel, using exponential expansion for the bath memory kernel.
2. Multiple rounds of MPO squaring and SVD compression (e.g., via iDMRG or variational canonicalization) to achieve exponentiation with minimal bond growth.
3. Multiplication and compression of infinite GMPS representations for the impurity propagator and bath influence functionals.
4. Contraction of observables using transfer-matrix methods and boundary eigenvectors, with costs that grow only with the number of measurement points (not total evolution time) (Sun et al., 2024).

4. Algorithmic Steps and Numerical Scaling

The algorithm reduces the computational complexity by exploiting the exponential decay of the hybridization kernel (as ensured by $|\lambda|<1$ in the Prony expansion) and the time-translational invariance at $T=0$:

• Each influence functional starts with bond dimension $\sim2n+2$, growing to $\chi_I\sim O(n)$ after exponentiation, independent of $N$.
• The full ADT bond dimension is at most $\chi_A \leq \chi_K \cdot \chi_I^M$ for $M$ impurity flavors.
• Contrary to finite-T TEMPO, which requires open-boundary MPS of length $N$ and sub-exponential scaling in $\chi$, the iGTEMPO construction achieves fixed resource usage as $\beta\rightarrow\infty$ (Guo et al., 2024).

Efficient implementations rely on least-squares or Padé-based exponential kernel fitting, iterative MPO squaring/truncation, and the use of infinite-MPS compression routines (e.g., MPSKit.jl), making iGTEMPO orders-of-magnitude faster than previous approaches for long-time or zero-temperature impurity solvers (Guo et al., 2024).

5. Benchmark Results and Comparative Performance

Extensive benchmarks demonstrate:

• Exact agreement for the Toulouse model (single flavor, non-interacting limit) between iGTEMPO and analytic solutions with $\delta\tau\sim0.01$-$0.1$, $\chi\sim30$–$60$, and $n\sim25$-$40$ over $G(\tau)$ up to $50/D$, with errors $\lesssim10^{-3}$ (Guo et al., 2024).
• For the single-orbital Anderson model and two-orbital variants, iGTEMPO matches CTQMC results to within 1–5% over all relevant imaginary times, with bond dimension requirements dramatically lower than for finite-$T$ GTEMPO (Guo et al., 2024, Chen et al., 2023).

For real-time and non-equilibrium problems, infinite Grassmann MPO methods retain cost independent of total evolution time. Comparison to conventional finite TEMPO reveals a typical speed-up of $10$–$100\times$ for long-time simulations, with the accuracy preserved for Green's functions in both equilibrium and transport setups (Guo et al., 2024, Sun et al., 2024).

6. Limitations, Applicability, and Future Directions

The primary limitation of iGTEMPO arises for large impurity flavor number $M$, where bond dimension scaling becomes prohibitive, and aggressive compression or symmetry exploitation is necessary. Additionally, applicability is currently limited to systems where the bath kernel is truly time-translational invariant; explicitly time-dependent driving or rapid bath variations require generalizations—sometimes involving windowed or non-uniform MPS strategies (Sun et al., 2024).

Recent studies indicate the potential for integration into dynamical mean-field theory (DMFT) loops, extension to multi-bath or multi-orbital problems, and hybridization with CTQMC or NRG techniques for more challenging parameter regimes (Guo et al., 2024, Chen et al., 2023).

7. Implementation Optimizations and Practical Recommendations

State-of-the-art iGTEMPO implementations incorporate:

• Prony or Padé decomposition for minimal-exponential kernel representation.
• Iterative MPO squaring with SVD-based truncation after each round.
• Infinite MPS compression and transfer-matrix diagonalization for long-time correlations.
• Analytical handling of bath integrals, eliminating discretization error.
• Direct evaluation of imaginary-frequency quantities via Fourier-diagonal operators.
• High parallelizability over measurement points and flexibility to post-process observables (Guo et al., 2024).

For practical deployment, the methodology is noise-free, sign-problem-free, and achieves computational costs $O(\chi^{2M})$ with $\chi\sim 10$–$50$, making it an optimal choice for moderate-sized zero-temperature quantum impurity problems.

---

In summary, the TEMPO family, and in particular its infinite Grassmann variant (iGTEMPO), constitutes a major advance in tensor-network-based quantum impurity solvers. By harnessing time-translational invariance and efficient MPO exponentiation and compression techniques, iGTEMPO enables numerically exact, highly scalable, and resource-efficient simulation of zero-temperature and steady-state quantum impurity models previously intractable for traditional quantum Monte Carlo or finite-memory tensor approaches (Guo et al., 2024, Sun et al., 2024, Guo et al., 2024, Chen et al., 2023, Guo et al., 2024).