Enhancing the ergodicity of Worldvolume HMC via embedding Generalized-thimble HMC

This paper proposes and validates a hybrid algorithm that embeds Generalized-thimble HMC within Worldvolume HMC to overcome ergodicity limitations in thin-layer simulations, thereby enabling efficient, large-scale studies of the doped Hubbard model with controlled statistical errors.

The Gist

The Big Problem: The "Oscillating Wave"

Imagine you are trying to calculate the total weight of a pile of sand. In most physics problems, every grain of sand has a positive weight, so you just add them up. But in certain quantum systems (like the Hubbard model, which describes how electrons move in a material), the "weight" of each configuration isn't just a number; it's a wave that oscillates between positive and negative values.

If you try to add up billions of these waves, the positive ones cancel out the negative ones almost perfectly. This is called the Sign Problem. It's like trying to hear a whisper in a hurricane; the signal is there, but the noise (the cancellations) makes it impossible to measure anything useful without an impossibly huge amount of data.

The Old Solution: The "Deformed Map"

To fix this, physicists use a trick called the Lefschetz Thimble method. Imagine the original problem is a flat, foggy map where the fog (the oscillations) is so thick you can't see anything. The solution is to lift the map into a 3D space and stretch it into a new shape (a "deformed surface"). On this new shape, the fog clears up, and the waves stop oscillating so wildly.

However, there's a catch. As you stretch the map to clear the fog, it can get torn into separate islands. If your computer simulation (a "walker") gets stuck on one island, it can't jump to the others because the gap is too wide. This is an Ergodicity Problem—the simulation gets stuck and stops exploring the whole picture.

The Current Best Tool: The "Worldvolume"

A method called Worldvolume Hybrid Monte Carlo (WV-HMC) was invented to solve the "stuck on an island" problem. Instead of staying on one specific shape, WV-HMC lets the simulation wander through a "worldvolume"—a continuous tunnel connecting all the different shapes (from the flat map to the fully stretched 3D shape).

Think of WV-HMC as a hiker walking through a valley that connects all the islands. This works great, but it has a limitation: if the "valley" is very narrow (a thin layer), the hiker moves very slowly and inefficiently. They keep bumping into the walls, and it takes forever to explore the area.

The New Innovation: The "Hybrid Hiker"

This paper proposes a new strategy: Embedding Generalized Thimble HMC (GT-HMC) into WV-HMC.

Here is the analogy:

• WV-HMC is like a hiker walking through a narrow, winding tunnel (the worldvolume). It's safe and connects everything, but the tunnel is so thin that the hiker has to take tiny, cautious steps.
• GT-HMC is like a hiker who is allowed to run freely on a specific, wide plateau (a single deformed surface). They can take huge, fast strides. However, if they run too far, they might fall off the edge of the plateau (ergodicity issues).

The Solution: The authors created a hybrid system.

1. Most of the time, the hiker walks through the narrow tunnel (WV-HMC) to ensure they don't get stuck on one island and can visit all the necessary areas.
2. Occasionally, the hiker steps out onto the wide plateau (GT-HMC) to take giant, efficient strides and cover ground quickly.

The paper proves mathematically that these two modes can be mixed together without breaking the rules of physics. The "tunnel" and the "plateau" are actually part of the same geometric structure, so switching between them is seamless.

Why This Matters for the Hubbard Model

The authors tested this on the doped Hubbard model (a model for high-temperature superconductors).

• They found a special "knob" (a parameter called $\alpha$) that they could turn. Turning this knob made the "fog" (the sign problem) disappear almost immediately, meaning they didn't need to stretch the map very far.
• Because they didn't need to stretch the map far, the "tunnel" (worldvolume) became very thin.
• A thin tunnel is usually bad for the standard WV-HMC method because it's too slow.
• The Result: By using their new hybrid method (WV-HMC + GT-HMC), they were able to simulate much larger systems on a computer than before. They successfully calculated the energy and particle density of the system with high precision, even though the "tunnel" was very thin.

Summary

The paper introduces a clever way to combine two different simulation techniques. It's like giving a slow, careful explorer a pair of running shoes for the open plains, while keeping their safety harness for the narrow bridges. This allows them to explore complex quantum systems faster and more accurately, specifically solving a problem where the simulation space becomes too narrow for standard methods to work efficiently.

Technical Summary

Technical Summary: Enhancing the Ergodicity of Worldvolume HMC via Embedding Generalized Thimble HMC

Problem Statement
The numerical sign problem remains a primary obstacle in first-principles computations of quantum many-body systems, such as finite-density QCD and strongly correlated electron systems. When the action becomes complex, the Boltzmann weight oscillates, leading to severe cancellations that render Monte Carlo sampling exponentially difficult. While the Lefschetz thimble method mitigates the sign problem by deforming the integration surface into the complexified space, it introduces a trade-off: deep deformations suppress oscillations but partition the surface via zeros of the Boltzmann weight, causing ergodicity issues where Markov chains cannot traverse between disconnected regions.

The Worldvolume Hybrid Monte Carlo (WV-HMC) method was developed to address this by treating the flow time as a dynamical variable, allowing sampling over a continuous union of deformed surfaces (the "worldvolume"). However, WV-HMC faces a specific limitation when applied to systems where the sign problem can be suppressed with a small maximum flow time (e.g., the doped Hubbard model using a redundant parameter $\alpha$). In such cases, the worldvolume compresses into a thin layer. Because WV-HMC performs isotropic momentum refreshment, phase-space exploration becomes inefficient in this thin geometry, causing ergodicity problems to reemerge.

Methodology
To resolve the ergodicity issues inherent in thin worldvolumes, the authors propose embedding the Generalized Thimble Hybrid Monte Carlo (GT-HMC) algorithm into the WV-HMC framework.

1. Theoretical Framework:

   • GT-HMC: Operates on a single deformed surface $\Sigma_t$ at a fixed flow time. While it suffers from ergodicity issues at the zeros of the Boltzmann weight, it allows for highly efficient exploration within allowed regions and permits larger molecular dynamics (MD) step sizes compared to WV-HMC because configurations feel repulsion from zeros only along the tangent direction.
   • WV-HMC: Operates on the $(N+1)$-dimensional worldvolume $\mathcal{R}$, a continuous union of surfaces $\Sigma_t$. It utilizes a symplectic integrator (RATTLE) to preserve the symplectic volume form without computing Jacobians.
   • Embedding: The core theoretical contribution is a rigorous proof that the tangent bundle $T\Sigma_t$ used in GT-HMC can be symplectically embedded into the tangent bundle $T\mathcal{R}$ used in WV-HMC. By parametrizing the momentum in the worldvolume to include a component $e$ along the flow direction, the authors show that restricting $e=0$ and $t=$ constant recovers the GT-HMC dynamics. This allows GT-HMC updates to be treated as valid subprocesses within the WV-HMC Markov chain, satisfying detailed balance and preserving the symplectic structure.

2. Algorithm Implementation:
   The combined algorithm alternates between "pure" WV-HMC updates (which explore the flow-time direction) and embedded GT-HMC updates (which explore the spatial configuration space at fixed flow times). The MD step sizes can be tuned independently for each subprocess, allowing GT-HMC to utilize larger steps where appropriate.

3. Application to the Hubbard Model:
   The method is applied to the two-dimensional doped Hubbard model on an $8 \times 8$ lattice. The authors utilize a redundant, nonphysical parameter $\alpha$ (introduced in prior work) to significantly reduce the sign problem on the original integration surface, thereby allowing the maximum flow time $T_1$ to be kept small.

   • Parameter Tuning: $\alpha$ is tuned to the smallest value that avoids ergodicity issues (avoiding long plateaus in phase factor histories). The upper cutoff $T_1$ is set to small values ($4.0 \times 10^{-3}$ to $8.0 \times 10^{-3}$) sufficient to suppress oscillations but small enough to create a thin worldvolume.
   • Observables: Number density and energy density are calculated and extrapolated to the zero Trotter step limit ($\epsilon \to 0$) at fixed temperature $T/t \approx 0.156$ and interaction $U/t = 8.0$.

Key Results

• Validation: The authors confirm that the standalone GT-HMC, pure WV-HMC, and the combined algorithm yield results consistent within statistical errors for the $8 \times 8$ lattice at Trotter step $\epsilon = 0.32$. This validates the mathematical embedding and the practical efficacy of the combined approach.
• Continuum Limit: Using the combined algorithm, the authors successfully simulated larger spacetime lattices (up to $N_t = 24$) and extrapolated observables to the continuum limit ($\epsilon \to 0$). The results show $O(\epsilon^2)$ scaling, consistent with theoretical expectations.
• Statistical Control: Even with modest sample sizes, the combined algorithm achieves controlled statistical errors across the entire range of the chemical potential. The extrapolated results agree with Auxiliary Field Quantum Monte Carlo (ALF) data at $\epsilon = 0.01$, resolving previous discrepancies observed in pure WV-HMC results which were attributed to finite-$\epsilon$ effects.

Significance and Claims
The paper claims that the proposed embedding of GT-HMC into WV-HMC provides a robust solution for simulating systems where the sign problem is mild enough to allow small flow times but where the resulting thin worldvolume hinders standard WV-HMC ergodicity. By combining the strengths of both methods—GT-HMC's efficiency in exploring allowed regions with larger step sizes and WV-HMC's ability to traverse the flow-time dimension—the authors enable simulations on larger spacetime lattices that were previously computationally prohibitive or ergodically trapped.

The authors note that while the current implementation uses direct solvers with $O(N^3)$ scaling, the embedding prescription is general and is expected to remain applicable to future implementations using pseudofermions and iterative solvers ($O(N^2)$ scaling), which would further reduce computational costs. The work demonstrates the feasibility of this approach for the doped Hubbard model, offering a pathway to study larger systems and approach the continuum limit with controlled errors.