Abstract
Tensor Network States (TNS) are powerful tools for simulating quantum many-body systems, yet their numerical stability is often compromised by the topological properties of the underlying network. This paper investigates the "closedness" of various TNS sets, proving that while loop-free architectures like Matrix Product States (MPS) with open boundaries are closed, architectures containing loops—such as PEPS and periodic MPS—are generally non-closed. This lack of closedness leads to numerical divergence in optimization algorithms as they attempt to reach limit points outside the representable set. We provide mathematical proofs for these properties and propose regularization techniques to mitigate instability in non-closed architectures.
Introduction: The Challenge of Numerical Stability
The simulation of quantum many-body systems is notoriously difficult due to the "curse of dimensionality," where the complexity of the Hilbert space grows exponentially with the number of particles. Tensor Network States (TNS) address this by approximating the system's state using a much smaller set of interconnected tensors.
A significant hurdle in TNS simulations is numerical instability, where tensor elements diverge toward infinity during optimization. This paper identifies the root cause as a topological deficiency: the set of states being optimized over is often not "closed." In topology, a set is closed if it contains all its boundary points. If a TNS set is not closed, an optimization algorithm may attempt to reach a "limit point" on the boundary that is not actually part of the set. As the algorithm approaches this non-existent point, the tensor elements must grow larger and larger to compensate, leading to the observed numerical divergence.
Methodology
The authors utilize a combination of rigorous mathematical proofs and constructive counter-examples to analyze the geometry of TNS sets:
- Geometric Analysis: TNS sets are mapped to well-known mathematical structures, such as Grassmann and Stiefel manifolds, to prove closedness for specific architectures.
- Canonical Forms: By utilizing "gauge freedom"—the ability to change tensor elements without altering the physical state—the authors bring tensors into standard forms that simplify the study of their topological properties.
- Regularization: To address the instability of non-closed sets, the paper proposes new mathematical "penalties" (regularization terms) for optimization algorithms to prevent tensor elements from diverging.
Benchmark States and Models
As a theoretical study, this work uses well-known quantum states and models as benchmarks rather than traditional datasets:
- The W State: A specific entangled state used to prove that periodic MPS sets are not closed.
- The AKLT State: Used as a case study for the bilinear-biquadratic spin-1 chain to demonstrate scenarios where optimizers exist within the interior of a set.
- Two-domain states: Newly defined states used to demonstrate the non-closedness of Projected Entangled-Pair States (PEPS).
Key Insights: The Role of Network Topology
The most striking insight is that the geometry of the network dictates its numerical stability.
- Loops are the Culprit: Architectures without loops (e.g., Tree Tensor Networks) or those with specific isometric constraints (e.g., MERA) are always closed. However, adding loops—as in Periodic Boundary Conditions (PBC) or 2D grids (PEPS)—generally renders the set non-closed.
- The "W State" Paradox: For systems with periodic boundaries, one can get arbitrarily close to a W state by increasing tensor values, but the state itself cannot be represented exactly with a fixed bond dimension. This "edge case" frequently breaks numerical solvers.
- Divergence as a Signal: When a simulation's tensor elements start to diverge, it is often a sign that the physical state the algorithm is seeking lies just outside the reachable manifold of the chosen tensor network.
Results and Implications
The paper categorizes common TNS architectures based on their stability:
- Proven Closed (Stable): Matrix Product States with Open Boundary Conditions (OBC-MPS), Tree Tensor Network States (TTNS), and the Multiscale Entanglement Renormalization Ansatz (MERA).
- Proven Non-Closed (Potentially Unstable): Translation-invariant and heterogeneous MPS with periodic boundaries, and Projected Entangled-Pair States (PEPS).
For translation-invariant periodic MPS, the authors provide specific bounds on the system size relative to the bond dimension where non-closedness is guaranteed. Furthermore, they demonstrate that adding a regularization term—specifically one that penalizes the norm of the transfer matrix—can effectively stabilize simulations that would otherwise diverge.
Limitations and Future Work
While this work provides a rigorous foundation, several questions remain:
- Tightness of Bounds: The mathematical bounds provided for non-closedness are "worst-case" scenarios; in practice, sets might behave as if they are closed for much smaller system sizes.
- Infinite Systems: A full topological treatment of scale-invariant MERA (siMERA) in the thermodynamic limit remains an open question.
- Undecidability: For complex 2D networks like iPEPS, it is mathematically "undecidable" whether a state even has a valid norm, suggesting that the boundary between stable and unstable simulations is even more complex than currently understood.
Summary updated