Tensor network manifolds and Riemannian fundamental theorem for tensor networks

This paper establishes a Riemannian fundamental theorem for various tensor network families by characterizing the interaction between their inherent gauge freedom and Riemannian manifold structure through the use of group actions and Riemannian submersions.

The Gist

Imagine you are trying to describe a massive, complex 3D sculpture. You could try to list the coordinates of every single atom, but that would take forever and be impossible to manage. Instead, you decide to build the sculpture out of smaller, manageable blocks (like LEGO bricks) that snap together in a specific pattern. This is essentially what Tensor Networks do for quantum physics: they break down incredibly complex, high-dimensional data (like the state of a quantum computer or a material) into a network of smaller, connected pieces.

However, there's a catch. Just like you could build the same LEGO castle using different colored bricks or by snapping the pieces together in a slightly different order, there are many different ways to arrange the "blocks" in a tensor network to represent the exact same final result. In math and physics, this is called gauge freedom. It's a bit of a nuisance because it means your map (the network) has extra, unnecessary details that don't change the destination (the physical state).

The Problem: Too Many Maps for One Destination

The paper tackles a specific problem: How do we get rid of these extra, redundant details so that every unique physical state has exactly one unique map?

The authors look at several different types of these "block networks" (like Matrix Product States, which are like a long chain of blocks, or PEPS, which are like a 2D grid of blocks). They want to find a rule that says, "If you change the blocks in this specific way, you haven't actually changed the sculpture; you've just rearranged the scaffolding."

The Solution: A Mathematical "Filter"

The authors use a branch of mathematics called Riemannian geometry. To use a simple analogy, imagine the space of all possible ways to build your LEGO sculpture is a giant, bumpy landscape.

• The Landscape (Manifold): Every point on this landscape is a different way you could arrange your blocks.
• The Redundancy (Gauge): Some points on this landscape look different but actually represent the exact same sculpture. They are like different paths leading to the same mountain peak.
• The Goal: The authors want to create a "quotient" landscape. This is a new, smoother map where all the redundant paths are squashed together. On this new map, every single point corresponds to exactly one unique sculpture, with no duplicates.

The "Riemannian Fundamental Theorem"

The paper's main achievement is proving that for several important types of tensor networks, you can indeed create this perfect, non-redundant map. They call this the Riemannian Fundamental Theorem.

Here is how they did it, using their own metaphors:

1. Identify the Symmetry: They figured out exactly how you can swap or rotate the "blocks" (tensors) without changing the final result. They found that these swaps act like a group action—think of it as a set of rules for how you can spin or flip your LEGO pieces.
2. The Smooth Slide: They proved that if you apply these rules, the landscape of possibilities behaves nicely. Specifically, they showed that the process of squashing the redundant paths together is a Riemannian submersion.
   • Analogy: Imagine a waterfall. The water flowing down represents all the different ways to build the network. The pool at the bottom represents the unique physical states. The authors proved that the water flows down smoothly and evenly, so that if you know where a drop of water ends up in the pool, you know exactly which "path" it took down the waterfall, up to the specific "twists" (gauge) that don't matter.

What They Studied

The paper doesn't just look at one type of network; they tested their "filter" on several common families used in quantum physics:

• 1D and 2D Quantum Circuits: Like a circuit board with layers of gates.
• Matrix Product States (MPS): A long chain of connected tensors (very common in 1D systems).
• Projected Entangled Pair States (PEPS): A 2D grid of tensors (used for 2D systems).
• Sequentially Generated States: States built up row by row.
• Isometric PEPS: A specific type of PEPS where the blocks have special "locking" properties.

The Takeaway

The paper claims that for all these families, we can now mathematically define a "perfect" space where:

1. Every point represents a unique quantum state.
2. There is no confusion or double-counting caused by the "gauge freedom" (the redundant ways to build the network).
3. This space is "smooth" and well-behaved, which means we can use powerful mathematical tools (like optimization algorithms) to navigate it efficiently.

In short, the authors have built a rigorous mathematical framework that cleans up the "messy" ways we describe quantum states, ensuring that when we try to optimize or analyze these systems, we are working with a clean, one-to-one map of reality. This is crucial for making the computer algorithms that simulate quantum matter more reliable and efficient.

Technical Summary

Technical Summary: Tensor Network Manifolds and Riemannian Fundamental Theorem for Tensor Networks

Problem Statement
Tensor networks serve as a powerful framework for representing high-dimensional data and many-body quantum states, often enabling efficient parametrizations where full Hilbert-space descriptions are computationally intractable. A central feature of these networks is "gauge freedom": distinct choices of local tensors can yield the same global tensor. While fundamental theorems characterizing this gauge freedom exist for specific families like Matrix Product States (MPS), a unified geometric understanding of gauge freedom across broader families of tensor networks has been lacking.

The paper addresses the need to formalize the interaction between the Riemannian manifold structure of tensor network parameter spaces and their inherent gauge freedom. Specifically, the authors aim to characterize the gauge freedom of several tensor network families not merely as an equivalence relation, but as a group action that induces a Riemannian submersion. This structure is crucial because it allows the application of Riemannian optimization and sampling algorithms (developed in prior works) with rigorous convergence guarantees to these specific tensor network families.

Methodology
The authors employ tools from differential geometry, specifically the theory of Lie groups and Riemannian submersions. The core methodology involves:

1. Defining Tensor Network Manifolds: The authors define tensor network families as smooth manifolds (specifically real manifolds, distinct from complex manifold approaches in some prior literature) constructed from products of spheres, unitary groups, and Stiefel manifolds, endowed with natural Riemannian metrics (e.g., round metrics, bi-invariant metrics, Fubini-Study metrics).
2. Characterizing Gauge Freedom: For each family, the authors derive "fundamental theorems" that explicitly describe the relationship between two tensor networks representing the same global state (up to a global phase). This involves proving that if two networks represent the same state, their local tensors are related by the multiplication of unitary matrices and their inverses over the contracting edges.
3. Constructing Group Actions: The identified gauge freedom is formalized as a smooth, free, and isometric action of a specific Lie group (typically products of unitary groups $U(d)$ or projective unitary groups $PU(d)$) on the tensor network manifold.
4. Establishing Riemannian Submersions: By proving the actions are free and isometric, the authors demonstrate that the quotient spaces (orbit spaces) admit smooth manifold structures and that the projection maps from the original manifold to the quotient are Riemannian submersions. This ensures the quotient space inherits a well-defined Riemannian metric.

Key Contributions and Results
The paper establishes a "Riemannian fundamental theorem" for the following tensor network families, proving that they define quotient tensor network manifolds where points in the quotient correspond one-to-one (or one-to-one up to a phase) with the physical states or circuits represented:

1. Depth-Two Quantum Circuits (1D and 2D):

   • With Fixed Input: The authors characterize the gauge for 1D and 2D depth-two circuits with fixed inputs. They show that the manifold of such circuits admits a quotient structure where points correspond to the circuit's output state exactly, and a separate quotient (involving projective spaces) where points correspond to the state up to a global phase.
   • Without Fixed Input: Similar results are derived for circuits without fixed inputs, where the entire set of gates constitutes the manifold.
   • Technical Note: For 2D circuits, the authors note that while a quotient manifold characterizing the state up to a phase exists, a quotient characterizing the state exactly (without phase ambiguity) via a free action is not guaranteed in the same manner as the 1D case.

2. Matrix Product States (MPS):

   • The paper reformulates known results for injective MPS in canonical form with open boundary conditions within the real manifold framework. It establishes quotient manifolds that uniquely characterize the state and the state up to a phase, confirming the gauge freedom is described by unitary multiplications on the virtual bonds.

3. Two-Dimensional Sequentially Generated States:

   • For states constructed by arranging MPS rows and stacking unitaries along columns, the authors identify the gauge freedom. They prove the existence of a quotient tensor network manifold that characterizes these states up to a global phase.

4. Isometric PEPS:

   • The authors study Isometric Projected Entangled Pair States (PEPS) on a lattice with a specific radial preparation order. They characterize the gauge freedom under full-rank assumptions on the isometries and the bottom states, establishing a quotient manifold structure for these states up to a phase.

Significance and Claims
The paper claims that its primary significance lies in providing a rigorous geometric framework for tensor networks that facilitates numerical optimization. By establishing that these families define quotient tensor network manifolds via Riemannian submersions, the work enables the direct application of advanced Riemannian optimization and sampling algorithms (cited from [10]) to these specific families.

The authors emphasize that the "Riemarian fundamental theorem" is not just a characterization of gauge symmetry but a structural guarantee that the orbit space is a smooth manifold with a compatible metric. This is presented as a necessary condition for importing powerful techniques from the rapidly growing field of Riemannian optimization (motivated by machine learning applications) to quantum many-body physics. The work generalizes the geometric framework introduced in [9] by shifting the perspective to real manifolds and explicitly constructing the quotient structures required for these optimization algorithms to function with rigorous convergence properties.

The paper does not propose new experimental setups or specific numerical benchmarks but rather lays the mathematical groundwork for future algorithmic developments by ensuring the underlying parameter spaces are well-behaved geometrically.