Benchmarking the accuracy of superconducting pair-pair correlations within Constrained Path Quantum Monte Carlo

This paper critically evaluates measurement techniques in Constrained Path Quantum Monte Carlo for the Hubbard model, finding that while the standard back propagation method underestimates superconducting pair-pair correlations, the more accurate constraint release technique offers a viable alternative despite its higher computational cost and reintroduction of the sign problem.

The Gist

Imagine you are trying to predict the weather in a tiny, chaotic city made of electrons. This city is governed by the Hubbard Model, a set of rules describing how these electrons interact, hop around, and sometimes team up to form a superconductor (a material that conducts electricity with zero resistance).

The problem is that this city is so complex and the electrons so numerous that you can't calculate the answer with a simple formula. You need a supercomputer simulation. This is where Quantum Monte Carlo (QMC) comes in. Think of QMC as sending out millions of tiny, random "scouts" (called walkers) to explore every possible path the electrons could take, then averaging their reports to find the truth.

The Big Problem: The "Sign" Trap

There's a catch. In the quantum world, these scouts can sometimes give conflicting reports that cancel each other out (positive and negative numbers canceling to zero). This is called the Fermion Sign Problem. If the city gets too big or the interactions too strong, the noise drowns out the signal, and your simulation crashes. It's like trying to hear a whisper in a hurricane.

The Usual Fix: The "Constrained Path"

To stop the hurricane, scientists use a method called Constrained Path Monte Carlo (CPMC). They give the scouts a map (a "trial wavefunction") and say, "Only walk on paths that look somewhat like this map." If a scout wanders off into a "negative" zone, you kick them out of the game. This keeps the simulation running, but it's an approximation. You are forcing the scouts to stay on a specific road, which might not be the exact road the electrons would take.

The Mystery: Measuring the "Super"

The main goal of this paper is to check if CPMC can accurately measure superconducting pair-pair correlations.

• The Analogy: Imagine you want to know how often two people in the city hold hands (pair up) to dance.
• The Issue: In the standard CPMC method, the "map" (trial wavefunction) is great for predicting the energy of the city, but it's bad at predicting who is holding hands because the act of measuring "hand-holding" changes the path the scouts took.

To fix this, scientists use a technique called Back Propagation (BP).

• How it works: Imagine the scouts walk forward to the future. To measure the hand-holding, BP makes them rewind their steps (back-propagate) to see where they started, using the saved map to guess the past.
• The Flaw: The paper finds that this "rewind" trick is flawed. It's like trying to reconstruct a messy room by looking at a blurry photo of it; you end up underestimating how messy it really was. The authors found that Back Propagation consistently underestimates how strongly electrons pair up. It tells you, "There's a little bit of dancing," when the reality might be "A huge dance party."

The Better (but Harder) Solution: "Constraint Release"

The authors tested a newer, more powerful technique called Constraint Release (CR).

• The Analogy: Instead of just rewinding the scouts with a blurry map, CR says, "Okay, let's let the scouts wander freely for a bit, but we'll use the best possible map we have to guide them, and we'll run a second, separate simulation just to double-check the hand-holding."
• The Result: This method is much more accurate. It correctly identifies the "dance party."
• The Catch: It is incredibly expensive computationally. It's like hiring a second team of 100 detectives just to verify the first team's notes. Also, because it lets the scouts wander a bit more freely, it risks running into the "Sign Problem" (the hurricane) again if the interactions get too strong.

What They Found

The team tested these methods on different "cities" (lattices):

1. One-Dimensional City (A straight line): The "rewind" trick (BP) worked perfectly here.
2. Ladders and Squares: As soon as the city got more complex (2D), the "rewind" trick started failing. It significantly underestimated the superconducting pairing.
3. The "Constraint Release" method: It gave the correct answer, proving that the "rewind" method was indeed lying to us by being too conservative.

The Takeaway

If previous studies used the "Back Propagation" method to claim that a material doesn't have superconductivity, those claims might be wrong. The method might have just been too weak to see the pairing that was actually there.

In short: The paper is a warning label for scientists. It says, "If you use the standard, fast way to measure electron pairing, you will likely miss the party. If you want to be sure, you need to use the slow, expensive, high-precision method, even if it's harder to run."

This suggests that we might need to re-evaluate many past studies on high-temperature superconductors, as they might have missed the very phenomenon they were looking for.

Technical Summary

1. Problem Statement

The Hubbard model is a fundamental theoretical framework for understanding unconventional superconductivity, particularly in high-$T_c$ cuprates and organic charge-transfer solids. While Quantum Monte Carlo (QMC) methods are powerful tools for studying strongly correlated systems, they are hindered by the Fermion sign problem, which causes an exponential loss of precision as system size, interaction strength ($U$), and inverse temperature ($\beta$) increase.

The Constrained Path Monte Carlo (CPMC) method mitigates the sign problem by imposing a constraint based on a trial wavefunction ($|\Psi_t\rangle$), ensuring the overlap between the walker and the trial wavefunction remains positive. While CPMC is highly accurate for calculating ground state energies, its accuracy for correlation functions—specifically superconducting (SC) pair-pair correlations—is less certain.

• The Core Issue: Calculating observables that do not commute with the Hamiltonian (like pair-pair correlations) requires an additional approximation known as Back Propagation (BP).
• The Question: Does the BP approximation accurately capture the magnitude of superconducting correlations, or does it introduce systematic errors? Furthermore, can the recently proposed Constraint Release (CR) technique provide more accurate results despite reintroducing the sign problem?

2. Methodology

The authors benchmarked CPMC against numerically exact methods (DMRG and extrapolated Path Integral Renormalization Group - PIRG) across various lattice geometries and interaction strengths.

Models and Observables

• Model: The Hubbard model with nearest-neighbor hopping ($t$) and onsite Coulomb interaction ($U$).
• Lattices Tested:
  • 1D chains (Open Boundary Conditions).
  • 2-leg ladders (Open Boundary Conditions).
  • Square lattices ($4\times4$ periodic).
  • Anisotropic triangular lattices ($6\times6$ periodic).
• Observable: Equal-time singlet pair-pair correlation $P(r)$ and the average long-range correlation $\bar{P}$.

Techniques Compared

1. Back Propagation (BP): The standard method where the trial wavefunction is propagated backward in imaginary time to estimate the expectation value of an operator. It relies on the constraint $\langle \Psi_t | \phi_i \rangle > 0$ being valid in reverse, which is not guaranteed.
2. Constraint Release (CR): A technique that relaxes the constraint during the measurement phase. It uses the saved Slater determinant of a random walker ($|\phi_i^0\rangle$) from the CPMC ensemble as the right-hand trial wavefunction ($|\Psi_R\rangle$) and the original trial wavefunction ($|\Psi_t\rangle$) as the left ($|\Psi_L\rangle$). This allows for a more accurate projection but reintroduces the sign problem, requiring Monte Carlo sampling over Hubbard-Stratonovich fields.
3. Trial Wavefunctions: The study utilized both simple non-interacting (free electron) wavefunctions and more complex QP-PIRG (Quantum Number projected Path Integral Renormalization Group) wavefunctions to test if improving $|\Psi_t\rangle$ mitigates BP errors.

3. Key Contributions

• Systematic Benchmarking: The paper provides a rigorous comparison of BP and CR against exact results (DMRG) and high-accuracy variational methods (QP-PIRG) across a wide range of parameters ($U$) and lattice geometries.
• Identification of BP Systematic Error: The authors demonstrate that BP systematically underestimates superconducting pair-pair correlations, particularly at larger interaction strengths ($U$) and longer projection times ($\beta$).
• Validation of Constraint Release: The study confirms that CR can yield essentially exact results for correlation functions, provided the sign problem does not become intractable.
• Optimization of CR Parameters: The paper investigates the optimal choice of the projection parameter $\tau$ in the CR algorithm, finding that $\tau \approx \beta/4$ often yields better convergence than the standard $\tau = \beta/2$.

4. Key Results

A. One-Dimensional Systems

• In 1D, the constrained path approximation is theoretically exact.
• Result: Both BP and exact DMRG results for pair-pair correlations agreed perfectly over three orders of magnitude. This suggests BP is exact in 1D, likely due to the specific sign properties of the 1D hopping term.

B. Two-Leg Ladders

• Result: BP significantly underestimated the long-range pair-pair correlation $P(r)$ by nearly an order of magnitude compared to DMRG.
• CR Performance: CR produced results much closer to the exact DMRG values.
• Limitation: For stronger interactions ($U=8$), the average sign in CR dropped too low to be computationally feasible, highlighting the trade-off between accuracy and the sign problem.

C. Square Lattice ($4\times4$)

• Energy vs. Correlations: While CPMC mixed energy estimates were highly accurate (errors < 1%), the BP estimates for pair-pair correlations showed significant deviations.
• $\beta$ Dependence: BP results were sensitive to the imaginary time step $\beta$. For large $\beta$, BP underestimated pairing by ~10%.
• Trial Function Impact: Improving the trial wavefunction (using QP-PIRG) improved energy estimates but did not significantly improve the BP correlation results for large $\beta$.
• CR Performance: CR converged to the exact value for both $U=4$ and $U=8$. The study found that using $\tau = \beta/4$ (projecting longer on the side of the walker determinant) yielded faster convergence than $\tau = \beta/2$.

D. Anisotropic Triangular Lattice ($6\times6$)

• Context: This lattice models organic superconductors. The ground state is expected to be non-superconducting at half-filling for certain parameters.
• Result: BP again underestimated the magnitude of $\bar{P}$ compared to extrapolated QP-PIRG results, with the discrepancy growing as $U$ increased.
• CR Performance: CR agreed very well with the extrapolated QP-PIRG results, confirming its accuracy even in frustrated geometries near the metal-insulator transition ($U \sim 5$).

5. Significance and Conclusions

• Re-evaluation of Past Literature: The finding that BP underestimates superconducting correlations suggests that previous CPMC studies claiming "absence of superconductivity" or weak pairing in Hubbard models may have been conservative. The true pairing correlations could be stronger than previously reported.
• Methodological Guidance:
  • BP is computationally efficient but unreliable for quantitative predictions of pairing correlations, especially at strong coupling.
  • CR is the preferred method for high-accuracy correlation measurements, provided the system size and interaction strength do not make the sign problem intractable.
  • Trial Functions: Simply improving the trial wavefunction does not fix the systematic errors inherent in the BP approximation for correlations.
• Future Directions: The authors suggest that optimizing CR parameters (like $\tau$ and the number of Monte Carlo sweeps) and potentially developing hybrid methods (releasing only a subset of fields) could make CR more accessible for larger systems.

In summary, the paper establishes that while CPMC is a robust tool for ground state energies, Back Propagation is insufficient for accurately quantifying superconducting pair-pair correlations. The Constraint Release technique is necessary for benchmarking and validating superconducting tendencies in strongly correlated materials, despite its higher computational cost.