Stabilizing the parquet problem

This paper analyzes the stability of iterative parquet equation solutions by identifying a new source of convergence failure unrelated to vertex divergences and proposes a controlled stabilization strategy that successfully recovers physical solutions in strongly interacting regimes.

The Gist

The Big Picture: The "Stuck" Calculator

Imagine you are trying to solve a very complex math puzzle to understand how electrons behave in a material. You have a specific recipe (an algorithm) called the Parquet equations to solve this.

Usually, you start with a guess, plug it into the recipe, get a new answer, and repeat the process. You hope that with every step, your answer gets closer and closer to the "true" physical reality. This is called a fixed-point iteration.

However, the authors of this paper discovered that when the interactions between electrons get very strong (the "strong-coupling" regime), this recipe often gets stuck. It doesn't stop working; it just starts converging to a wrong answer. It's like a GPS that confidently tells you to drive into a lake because it got confused by a complex intersection. The computer thinks it has found the solution, but it's actually a "misleading convergence" to a fake reality.

The Culprit: The "Jacobian" Map

To figure out why the recipe gets stuck, the authors looked at the Jacobian. Think of the Jacobian as a topographic map of the solution landscape.

• Stable Ground: If you are on a gentle slope, and you take a step, you naturally roll back toward the bottom (the correct answer).
• Unstable Ground: Sometimes, the landscape has a "hill" or a "cliff" right where the correct answer should be. If you are there, even a tiny nudge sends you rolling away to a different valley (the wrong answer).

The paper found that in strong interactions, the "correct" answer sits on a hill. The standard method (damped iteration) tries to slow you down (damping) to keep you from rolling off, but sometimes the hill is so steep that slowing down isn't enough. You still roll off the cliff.

The Discovery: It's Not Just One Thing

Previously, scientists thought the recipe only broke when a specific mathematical "singularity" (a vertex divergence) appeared. They thought, "If we see this spike, the method will fail."

The authors proved this is not true.

• The Analogy: Imagine a car engine that stalls. Everyone thought it only stalled when the fuel line got clogged (vertex divergence). But the authors found the engine also stalls when the spark plugs are just slightly misaligned, even if the fuel line is perfectly clear.
• The Result: The method can fail before the big spikes appear, simply because the mathematical landscape has turned into a hill that pushes the solution away.

The Solution: The "Anti-Gravity" Stabilizer

The authors invented a Stabilization Strategy.

Imagine you are trying to balance a broom on your hand.

1. Standard Method: You just move your hand to keep the broom upright. If the broom starts falling too fast, you can't catch it.
2. The New Method: The authors realized that the broom is falling because of a specific direction (e.g., it's tipping to the left). Instead of just moving your hand, they put a tiny, invisible magnet on the broom that pushes it back toward the center only when it starts to tip in that specific dangerous direction.

Technically, they analyzed the "map" (the Jacobian), found the specific directions where the solution is unstable, and flipped the sign of the correction in those directions.

• If the math says "move forward," but that direction is unstable, the new method says "move backward."
• This turns the "hill" back into a "valley," allowing the calculation to roll back to the correct physical answer, even in very strong interactions.

The Proof: Two Simple Models

To prove this works, they tested it on two simplified "toy" models:

1. The Zero-Point Model: A very simple, abstract model with no spatial complexity.
2. The Hubbard Atom: A model representing a single atom where electrons repel each other strongly.

In both cases, the standard method failed and gave wrong answers once the interaction got strong. The new Stabilization Method successfully navigated through the "hills" and "cliffs," finding the correct physical solution even deep in the non-perturbative (very strong) regime.

A Twist: The "Strong-Coupling" Iteration

The paper also tried a different approach: instead of solving for the "parts" of the puzzle (reducible vertices), they solved for the "whole picture" (the full vertex).

• The Result: This approach had the opposite problem. It worked great when interactions were strong but failed when interactions were weak.
• The Metaphor: It's like a pair of shoes. One shoe fits perfectly when your foot is small (weak coupling) but falls off when your foot is big. The other shoe fits perfectly when your foot is big but slips off when your foot is small. The authors showed that by combining their stabilization trick with this "whole picture" approach, they could potentially cover all bases.

Summary

• The Problem: Standard methods for calculating electron behavior often fail in strong interactions, getting stuck on "wrong" answers that look like they are converging.
• The Cause: The mathematical landscape becomes unstable (like a hill) in specific directions, not just when obvious "spikes" appear.
• The Fix: A new algorithm that detects these unstable directions and flips the correction sign to push the solution back to the correct path.
• The Outcome: They successfully stabilized the solution for complex models where it previously failed, proving that the "wrong" answers were just a symptom of an unstable calculation, not a lack of a physical solution.

Technical Summary

Technical Summary: Stabilizing the Parquet Problem

Problem Statement
Self-consistent diagrammatic methods, particularly the parquet equations, are powerful tools for describing strongly correlated electron systems where charge, spin, and orbital fluctuations interplay. However, these iterative schemes frequently suffer from "misleading convergence," where the algorithm converges to a mathematically valid but unphysical fixed point, or fails to converge entirely. Historically, this instability has been attributed to the divergence of the two-particle irreducible vertex ($\Gamma$), which signals a branching of the Luttinger–Ward functional (LWF) and the breakdown of the skeleton expansion.

This paper identifies a critical gap in the current understanding: misleading convergence in parquet calculations is not exclusively caused by vertex divergences. The authors demonstrate that the physical fixed point of the parquet iteration can become unstable in parameter regimes where the irreducible vertex remains finite. Furthermore, standard damping techniques (mixing the current and previous iteration steps) are often insufficient to stabilize the solution in these regimes, particularly when the Jacobian of the iteration possesses eigenvalues with negative real parts.

Methodology
The authors analyze the stability of the iterative parquet scheme by examining the spectrum of the Jacobian matrix associated with the damped fixed-point iteration. The iteration is defined as $\Psi^{(n+1)} = p \mathcal{F}[\Psi^{(n)}] + (1-p)\Psi^{(n)}$, where $\Psi$ represents the set of reducible vertices and the Green's function, and $p$ is the damping parameter.

1. Jacobian Analysis: The stability of the fixed point is determined by the eigenvalues of the Jacobian $J = 1 - p\Pi$. The fixed point is unstable if any eigenvalue of $J$ has a magnitude greater than 1. The authors categorize the origins of instability into three cases based on the behavior of the eigenvalues of $\Pi$:

   • Case (i): An eigenvalue of $\Pi$ features an odd pole (associated with vertex divergences).
   • Case (ii): A real eigenvalue of $\Pi$ crosses zero.
   • Case (iii): The real part of a complex conjugate pair of eigenvalues crosses zero.
     Crucially, Cases (ii) and (iii) can lead to instability even in the absence of vertex divergences.

2. Stabilization Strategy: To address instabilities where standard damping fails (specifically when eigenvalues have negative real parts), the authors adapt a strategy previously proposed for self-consistent perturbation theory. Instead of using a scalar damping parameter $p$, they introduce a stabilization matrix $P$. This matrix is constructed by diagonalizing the Jacobian, identifying eigendirections corresponding to unstable eigenvalues (where the real part is $\le 0$), and flipping the sign of the damping parameter specifically for those directions. Mathematically, $P = U D U^{-1}$, where $U$ is the eigenbasis of the Jacobian and $D$ is a diagonal matrix containing $-p$ for unstable modes and $p$ for stable ones. This procedure locally stabilizes the physical fixed point without altering the fixed point itself.

3. Strong-Coupling Iteration: As an alternative approach, the authors propose an iteration scheme where the full vertex $F$ is iterated rather than the reducible vertices $\Phi$. This "strong-coupling iteration" inverts the Bethe-Salpeter equations to express the irreducible vertex as a functional of $F$. They find that this scheme exhibits inverted stability properties: the physical fixed point is unstable at weak coupling but becomes stable at strong coupling.

Key Results
The methodology is tested on two analytically solvable models: the Zero-Point (ZP) model and the Hubbard Atom (HA).

• Zero-Point Model: The stability phase space reveals that while the onset of instability at particle-hole symmetry coincides with a vertex divergence, instabilities occur at lower interaction strengths away from symmetry. These off-symmetry instabilities are driven by eigenvalues with zero real parts but finite imaginary parts (Case iii), occurring without any vertex divergence. The stabilization method successfully converges to the physical solution across these regions, whereas standard damped iteration fails.
• Hubbard Atom: In the HA, the number of unstable eigenvalues scales with the size of the fermionic frequency box ($N_f$). At half-filling, the first instability coincides with the first vertex divergence, but subsequent instabilities (including those driven by complex eigenvalues) appear at lower interactions or in different channels. The stabilization method allows convergence deep into the non-perturbative regime, crossing multiple divergence lines where standard methods converge to unphysical solutions or diverge.
• Comparison with Alternatives: The authors compare their stabilization method with the Newton-Raphson method, the Chord method (a quasi-Newton approach), and Anderson acceleration. While the Chord method converges faster, it shares the Newton-Raphson drawback that all fixed points (physical and unphysical) are locally stable, making the outcome highly sensitive to the initial guess. The stabilization method is shown to be more robust in ensuring convergence to the physical solution, even when the physical and unphysical fixed points cross.

Significance and Claims
The paper claims to provide a systematic explanation for the "misleading convergence" observed in parquet calculations, demonstrating that it is a consequence of the local instability of the physical fixed point, which can arise independently of vertex divergences.

• Decoupling from LWF Branching: Unlike self-consistent perturbation theory, where misleading convergence is linked to the multivaluedness of the LWF, the parquet formalism (which relies on the Schwinger-Dyson equation rather than a functional derivative of the LWF) suffers from instability solely due to the Jacobian spectrum. The divergence of the irreducible vertex is shown to be a sufficient but not necessary condition for this instability.
• Practical Stabilization: The proposed stabilization method offers a controlled strategy to recover the physical solution in regimes where standard damping fails. It allows for the calculation of parquet equations deep in the non-perturbative regime, even across multiple divergence lines.
• Future Implementation: The authors acknowledge that the computational cost of calculating and inverting the full Jacobian is prohibitive for realistic lattice models with momentum dependence. They suggest that future implementations will require vertex compression techniques (e.g., quantics tensor trains) or the use of predictors to approximate the Jacobian and initial guesses. They also note that the strong-coupling iteration scheme offers a complementary approach that may be stable in regimes where the conventional scheme is not.

In conclusion, the work establishes that the stability of the parquet iteration is governed by the spectral properties of its Jacobian and provides a rigorous, Jacobian-based algorithm to stabilize the physical solution, extending the applicability of parquet methods to strongly correlated regimes previously inaccessible due to convergence issues.