Symmetries and overparametrization properties of Hamiltonian variational ansatzes for the $(1+1)$d $\mathbb{Z}_2$ lattice gauge theory

This paper investigates five symmetry-preserving Hamiltonian variational ansatzes for the $(1+1)$d $\mathbb{Z}_2$ lattice gauge theory, demonstrating through numerical analysis of dynamical Lie algebras and quantum Fisher information matrices that overparametrization eliminates local minima and accelerates VQE convergence, thereby advancing the theoretical understanding of scalable quantum circuit design.

The Gist

Imagine you are trying to find the lowest point in a vast, foggy mountain range. This is what scientists call an optimization problem. In the world of quantum computing, they use a special tool called a Variational Quantum Algorithm (VQA) to do this. Think of the VQA as a hiker with a map that has adjustable knobs. Every time the hiker turns a knob, the map changes slightly, and they check if they are lower down the mountain. If they are, they keep going; if not, they try a different direction.

The "map" in this paper is called an Ansatz. It's a specific recipe for how the quantum computer builds its state. The authors of this paper studied five different recipes (labeled A through E) designed for a specific physics problem: the 1D Z2 Lattice Gauge Theory. You can think of this theory as a grid of tiny magnets and particles interacting with each other, governed by strict rules (symmetries) that nature follows.

Here is what the paper discovered, explained simply:

1. The "Over-Parameterized" Magic

Usually, when you have a mountain range with many knobs to turn, the hiker gets stuck in a small valley (a "local minimum") and thinks it's the bottom, even though a much deeper valley exists nearby. This is a common problem in quantum computing.

The paper found that if you give the hiker enough knobs (parameters), the small valleys disappear. The landscape becomes smooth, and the hiker can slide straight to the true bottom (the global minimum). This state is called overparameterization.

• The Analogy: Imagine trying to fold a piece of paper into a specific shape. If you only have a few folds, you might get stuck in a messy crumple. But if you have enough folds to make every tiny crease, you can perfectly achieve the shape without getting stuck.

2. The "Lie Algebra" and the "Search Space"

The authors wanted to know exactly how many knobs are needed before the small valleys disappear. To figure this out, they looked at two mathematical tools:

• The Dynamical Lie Algebra (DLA): Think of this as a list of all the possible directions the hiker can move. If the list is short, the hiker is stuck in a small room. If the list is long, the hiker can explore the whole mountain.
• The Quantum Fisher Information Matrix (QFIM): This measures how "flexible" the map is. When the rank of this matrix "saturates" (stops growing), it means the map has reached its maximum flexibility.

The paper showed that for their specific recipes, once the number of knobs exceeded a certain critical number, the QFIM stopped growing, and the "local valleys" vanished. The hiker could finally find the true bottom.

3. The "Three-Body" Twist

Most previous studies looked at simple interactions (like two magnets touching). This paper looked at a more complex interaction where three things interact at once (like three magnets influencing each other simultaneously).

• The Finding: Even with these complex three-way interactions, the "overparameterization" rule still held true. If you add enough knobs, the optimization problem becomes easy again.

4. The Speed of the Hiker

The authors also watched how fast the hiker moved down the mountain as they added more knobs.

• The Discovery: They found that the speed at which the hiker improved (the "decay rate" of the error) increased linearly with the number of knobs.
• The Analogy: It's like adding more engines to a car. The more engines you add, the faster the car goes, in a straight, predictable line. It doesn't suddenly jump to super-speed; it just gets steadily faster.

5. Not All Recipes Are Equal

The paper tested five different recipes (A, B, C, D, E).

• Recipes A, B, and C: These were "maximally expressive." They could explore every possible corner of the mountain.
• Recipe D: This one was limited. Even with many knobs, it couldn't reach the absolute bottom of the mountain because its "map" was missing certain directions.
• Recipe E: This was a special case. It had a very simple structure that scaled efficiently, suggesting it might be a good candidate for larger, more complex problems in the future.

Summary

In short, this paper is a guidebook for quantum computer designers. It proves that if you build your quantum "map" (ansatz) with enough adjustable knobs, you can avoid getting stuck in bad solutions. It also shows that the speed of finding the solution gets faster as you add more knobs, and that this works even for complex physics problems involving three-way interactions. The key takeaway is: More knobs (parameters) = Smoother path to the solution.

Technical Summary

Technical Summary: Symmetries and Overparametrization Properties of Hamiltonian Variational Ansatzes for the (1 + 1)d Z2 Lattice Gauge Theory

Problem Statement
Variational Quantum Algorithms (VQAs), particularly the Variational Quantum Eigensolver (VQE), face significant challenges in scalability, notably the "barren plateau" phenomenon where gradients vanish exponentially with system size, and the presence of spurious local minima that hinder convergence to global optima. Recent literature suggests that "overparametrization"—where the number of trainable parameters exceeds a critical threshold—can eliminate local minima and ensure convergence. However, the theoretical mechanisms governing overparametrization, specifically the relationship between the Dynamical Lie Algebra (DLA), the Quantum Fisher Information Matrix (QFIM), and the loss landscape, remain underexplored for ansatzes involving higher-weight Pauli strings (weight > 2) and complex symmetry structures. This is particularly relevant for Lattice Gauge Theories (LGTs), where Hamiltonians naturally involve multi-body interactions and possess both local (Gauss's law) and global symmetries that segment the Hilbert space.

Methodology
The authors investigate five distinct families of Hamiltonian Variational Ansatzes (HVA) derived from the Hamiltonian of a one-dimensional $Z_2$ lattice gauge theory (LGT) with spinless matter. The study proceeds through the following steps:

1. Model Definition: The system is defined on a chain of $N_s$ sites with periodic boundary conditions. The Hamiltonian includes hopping ($H_{hop}$), mass ($H_m$), and gauge ($H_g$) terms, involving two-body and three-body Pauli interactions. The model possesses local gauge symmetry (Gauss's law) and global symmetries including parity ($P$), charge conjugation ($C$), and total charge ($Q$).
2. Ansatz Construction: Five HVA families (labeled A through E) are constructed. Each family uses different subsets of the Hamiltonian terms as generators for the unitary circuit layers. The ansatzes are designed to respect specific subsets of the global symmetries, thereby restricting the evolution to specific invariant subspaces of the Hilbert space.
   • Families A, D, and E preserve $P$ and $C$.
   • Families B and C break $P$ and $C$ but preserve a combined symmetry $CP$.
   • Family E additionally preserves the gauge field inversion symmetry $Z$.
3. Invariant Subspace Analysis: The initial state $|\psi_0\rangle$ is chosen to be an eigenstate of the symmetry operators. The study focuses on the invariant subspace $S_X$ reachable by each ansatz family $X$. The dimension $d_X$ of these subspaces is calculated numerically.
4. DLA and QFIM Characterization:
   • The Dynamical Lie Algebra (DLA) $\mathfrak{g}_{S_X}$ is computed by taking the Lie closure of the projected generators within the invariant subspace. The dimension of this algebra, $\dim(\mathfrak{g}_{S_X})$, is determined.
   • The Quantum Fisher Information Matrix (QFIM) rank, $R^{(L)}_X$, is numerically evaluated for varying circuit depths $L$. The saturation value $R_X$ is identified as the maximum rank achievable.
5. VQE Optimization: The ansatzes are applied to find the ground state energy of the original Hamiltonian using VQE with gradient descent (GD) optimization. The study tracks the asymptotic loss (final energy) and the decay rate of the loss function ($\kappa$) as a function of the number of parameters.

Key Contributions and Results

• DLA Structure and Expressibility:

  • Families A, B, and C: These families exhibit DLA subrepresentations isomorphic to $\mathfrak{u}(d_X)$ or $\mathfrak{su}(d_X)$, where $d_X$ is the dimension of the invariant subspace. This indicates they are maximally expressible, capable of generating any unitary operation within their respective subspaces given sufficient depth.
  • Family D: The DLA is isomorphic to $\mathfrak{so}(d_D)$. This algebra is a proper subalgebra of $\mathfrak{su}(d_D)$, meaning the ansatz is not maximally expressible. Consequently, the ground state of the physical model may not lie within the orbit of the initial state, leading to potential failure in finding the true ground state even when overparametrized.
  • Family E: This family presents a unique case where the DLA is isomorphic to a direct sum of multiple copies of $\mathfrak{su}(2)$. Its dimension scales linearly with the system size $N_s$, contrasting with the exponential scaling of the other families.

• Overparametrization and QFIM Saturation:

  • The study confirms that the QFIM rank saturates at a critical number of layers $L_c$.
  • For maximally expressible families (A, B, C), the saturation rank $R_X$ equals $2d_X - 2$, consistent with the real degrees of freedom of the state vector.
  • For Family D, $R_D = d_D - 1$.
  • For Family E, the saturation rank is tightly bounded by $\dim(\mathfrak{g}_{S_E})$, which is significantly smaller than $2d_E - 2$.
  • The results support the theoretical bound that the QFIM rank is upper-bounded by the dimension of the DLA subrepresentation.

• Loss Landscape and Convergence:

  • Disappearance of Local Minima: As the number of parameters exceeds the critical threshold ($M > M_c$ or $L > L_c$), the variance of the asymptotic loss values across random initializations vanishes. This coincides with the saturation of the QFIM rank, confirming that overparametrization eliminates local minima in the loss landscape.
  • Ground State Recovery: While overparametrization removes local minima, it does not guarantee finding the true ground state if the ansatz is not expressible enough (e.g., Family D fails to reach the true ground state energy despite overparametrization).
  • Decay Rate Scaling: The decay rate of the loss function under gradient descent, $\bar{\kappa}$, scales linearly with the number of parameters (and thus with the number of layers $L$). Unlike the asymptotic loss, the decay rate does not exhibit an abrupt transition at the overparametrization threshold; it increases steadily through both under- and overparametrized regimes.

Significance and Claims
The paper claims to enrich the theory of quantum circuit overparametrization by extending the analysis to Hamiltonian variational ansatzes involving three-body interactions (weight-three Paulis) and complex symmetry sectors typical of lattice gauge theories.

The authors assert that their findings provide crucial insights for designing scalable VQAs:

1. Symmetry and DLA: The choice of generators and the resulting symmetries directly dictate the structure of the DLA, which in turn bounds the QFIM rank and the expressibility of the ansatz.
2. Overparametrization Criteria: Overparametrization is defined by the saturation of the QFIM rank, which is bounded by the DLA dimension. However, for systems with high expressibility, the bound derived from the real degrees of freedom of the state vector ($2d-2$) can be tighter than the DLA dimension bound.
3. Scalability Implications: The linear scaling of the DLA dimension in Family E suggests that it is possible to construct ansatzes that are overparametrizable with polynomial circuit depth while still maintaining the ability to converge to the ground state. This offers a potential pathway for scalable VQE algorithms in LGT simulations.

The authors remain modest regarding immediate scalability, noting that the current VQE setup with random initialization is suitable for toy models. They suggest that their results inform the design of more advanced, potentially scalable algorithms like ADAPT-VQE and SC-ADAPT-VQE, and identify future directions in analytically understanding generator choices and extending the framework to higher dimensions and non-Abelian LGTs.