Scaling up the transcorrelated density matrix renormalization group

This paper presents improved techniques for the transcorrelated density matrix renormalization group (DMRG) method, including optimized matrix product operators, entanglement-aware representations, and parameter tuning, which collectively enable large-scale calculations on $12 \times 12$ lattices and significantly reduce ground-state energy errors compared to standard DMRG.

The Gist

Imagine you are trying to solve a massive, incredibly complex puzzle. This puzzle represents the behavior of electrons in a material (like a metal or a superconductor). The more electrons you have, the more pieces the puzzle has, and the number of possible ways those pieces can fit together grows so fast that it becomes impossible for even the world's fastest supercomputers to solve it exactly.

This is the problem physicists face with strongly correlated systems. To get around this, they use shortcuts. One popular shortcut is called DMRG (Density Matrix Renormalization Group). Think of DMRG as a very smart, efficient way to look at the puzzle by focusing on the most important connections between pieces, ignoring the tiny, irrelevant details. It works great for 1D chains (like a line of beads), but when you try to use it on a 2D grid (like a chessboard), the "connections" get messy and long, making the puzzle hard to solve accurately without using a huge amount of computer memory.

Another technique, called the Transcorrelated Method, tries to fix the puzzle before you start solving it. Imagine the electrons are like people at a crowded party who are constantly bumping into each other. The "Transcorrelated Method" is like giving everyone a special "personal space bubble" (a mathematical tool called a correlator) that pushes them apart just enough so they don't bump into each other as much. This makes the party (the quantum system) much calmer and easier to manage.

However, there's a catch: creating these "personal space bubbles" makes the rules of the game (the Hamiltonian) much more complicated. It turns a simple set of instructions into a massive, tangled web of rules that is hard to feed into the DMRG solver.

What did this paper do?
The authors, Benjamin Corbett and Akimasa Miyake, figured out how to combine these two powerful methods (the "calming bubbles" and the "smart solver") to solve much larger puzzles than ever before. They successfully simulated a 2D grid of electrons up to 12x12 (144 sites), which is four times larger than previous attempts using this specific combination.

Here is how they did it, explained with three simple analogies:

1. The "Smart Filing System" (Optimized MPO Construction)

When you apply the "personal space bubbles" to the electrons, the rules of the game explode in complexity. It's like trying to organize a library where every book suddenly has 100 new chapters attached to it. If you try to organize this mess with a standard filing system, your computer crashes.

The authors invented a super-efficient filing system (a new algorithm for building "Matrix Product Operators"). Instead of trying to list every single rule individually, they found a pattern. They realized that even though the rules look messy, they are actually very sparse (mostly empty space) and have a hidden structure. By organizing the rules into a compact, block-diagonal format, they could feed a massive amount of information into the solver without breaking the computer. This allowed them to handle systems with billions of terms that were previously impossible.

2. The "Re-arranged Seating Chart" (Entanglement Mappings)

DMRG works best when the puzzle pieces are arranged in a line where neighbors are actually neighbors. But in a 2D grid, if you just read the grid row-by-row (like reading a book), you might jump from the left side of the room to the right side, creating a "long-distance connection" that confuses the solver.

The authors realized that the electrons in this specific "party" have a special seating preference.

• For sparse parties (few electrons): They arranged the seats based on the "energy shells" (like seating guests by how much energy they have). This kept the most active guests next to each other.
• For crowded parties (half-filling): They noticed that electrons on opposite sides of the room (like a checkerboard pattern) were actually best friends. So, they created a checkerboard seating chart (the "bipartite mapping"), pairing up opposite seats so that best friends sat right next to each other in the solver's line.

This re-arrangement meant the solver didn't have to reach across the room to understand the connections, making the solution much more accurate.

3. The "Self-Correcting Tuning Knob" (Optimizing the Correlator)

The "personal space bubble" has a dial (a parameter called J) that controls how big the bubble is.

• If the bubble is too small, the electrons still bump into each other.
• If the bubble is too big, the rules become too weird and the solution breaks.

In the past, scientists had to guess the right setting for this dial, or they would sometimes get a "fake" answer that looked great but was actually wrong (mathematically, it was "non-variational," meaning it could dip below the true ground state energy, which is physically impossible).

The authors created a self-correcting loop. They didn't just guess the dial setting; they adjusted the dial while the computer was solving the puzzle. They used a specific mathematical test (checking the "residual") to find the perfect setting that made the solution stable and accurate. This ensured they never got a "fake" low energy and always found the best possible answer for the size of the computer they were using.

The Result: A Bigger, Better Picture

By combining these three tricks, the authors were able to calculate the energy of these electron systems with 2.4 to 14 times more accuracy than standard methods using the same amount of computer power.

• The "Closed-Shell" Win: For systems where the electrons fill up the energy levels perfectly (like a full parking lot), the method was incredibly accurate, reducing errors by a factor of 14.
• The "Open-Shell" Challenge: For systems with "loose" electrons (like a half-full parking lot), the method still improved accuracy significantly (2.4x), though the electrons' chaotic nature made it harder to get perfect results.

In Summary:
This paper is like upgrading a GPS system. The old GPS (standard DMRG) got lost in complex 2D terrain. The "Transcorrelated" method was a new map that simplified the terrain but was too detailed to read. These authors built a better map reader (optimized filing), re-drew the roads to make sense of the terrain (smart seating), and added a self-correcting compass (tuning knob). The result? They can now navigate much larger and more complex quantum territories than ever before, bringing us closer to understanding how materials like superconductors work.

Technical Summary

1. Problem Statement

Calculating the ground-state properties of strongly correlated electron systems, such as the two-dimensional (2D) Fermi-Hubbard model, is a central challenge in condensed matter physics.

• The Challenge: The Hilbert space grows exponentially with system size, making exact solutions impossible for large systems. While the Density Matrix Renormalization Group (DMRG) is highly effective for 1D systems, it struggles with 2D systems due to long-range entanglement when mapped to a 1D Matrix Product State (MPS), requiring prohibitively large bond dimensions ($m$) for accuracy.
• The Transcorrelated (TC) Method: This method shifts a many-body correlator (e.g., Jastrow or Gutzwiller) from the wavefunction to the Hamiltonian via a similarity transformation ($\hat{H}_{TC} = e^{-\hat{\tau}} \hat{H} e^{\hat{\tau}}$). This produces an effective Hamiltonian where the ground state is less entangled and easier to approximate. However, the TC Hamiltonian is non-Hermitian (losing the variational principle) and introduces complex three-body terms, making its representation as a Matrix Product Operator (MPO) computationally expensive and limiting previous studies to small system sizes (e.g., $6\times6$ or $36$ sites).

2. Methodology

The authors apply the TC method to the momentum-space representation of the 2D Fermi-Hubbard model using a Gutzwiller correlator ($\hat{\tau} = J\hat{D}$, where $\hat{D}$ is the double-occupancy operator). The ground state $|\phi\rangle$ is represented as an MPS and optimized using DMRG. The study relies on three specific technical innovations to scale this approach:

A. Optimized MPO Construction

• Challenge: The TC Hamiltonian contains $O(N^5)$ terms for an $N$-site system. Standard MPO construction algorithms (like bipartite graph methods) result in bond dimensions scaling as $O(N^2)$, which is too expensive for large systems.
• Solution: The authors developed a hybrid MPO construction algorithm. This combines the rank-decomposition approach (which guarantees optimal bond dimension) with the bipartite graph method (which limits the number of operators to be proportional to the incoming bond dimension).
• Result: This reduces the MPO bond dimension from quadratic ($O(N^2)$) to linear ($O(N)$). Additionally, they implemented a novel algorithm to identify connected components in the operator graph, creating MPOs with optimal block-diagonal sparsity (up to 99.96% sparsity for $12\times12$ systems) without requiring prior knowledge of the system's symmetries.

B. Entanglement-Adaptive Mappings

• Challenge: Standard mappings (e.g., row-major) fail to capture the specific entanglement structure of momentum-space Hamiltonians, leading to high bond dimensions.
• Solution: The authors analyzed the entanglement structure of the ground state and proposed two new mappings:
  1. $\epsilon$-mapping (for dilute systems): Maps momentum modes to MPS tensors based on the decreasing value of the single-particle energy $\epsilon(k)$. This keeps correlations within energy shells "local" in the MPS.
  2. Bipartite mapping (for half-filling): Exploits the strong correlation between momentum modes $k$ and $k+\pi$ (which exist due to the bipartite nature of the lattice). It maps these pairs to adjacent tensors, significantly reducing the effective correlation length for half-filled systems.
• Result: These mappings outperform previous heuristic methods (like Fiedler optimization) and allow for more accurate DMRG calculations at lower bond dimensions.

C. Self-Consistent Correlator Optimization

• Challenge: The TC method is non-variational; the energy can drop below the true ground state, and the optimal correlator parameter $J$ is unknown. Previous works used fixed $J$ or optimized against a simple reference state.
• Solution: The authors optimize the parameter $J$ self-consistently alongside the MPS $|\phi\rangle$ at each bond dimension. They utilize two criteria:
  1. Minimizing the variance $\sigma^2 = \langle \phi | \hat{H}^\dagger \hat{H} | \phi \rangle - |\langle \phi | \hat{H} | \phi \rangle|^2$.
  2. Zeroing the residual $r(J) = \langle \phi | \hat{D} \hat{H} | \phi \rangle - \langle \phi | \hat{D} | \phi \rangle \langle \phi | \hat{H} | \phi \rangle$.
• Result: This ensures the calculated energy never falls below the reference ground state energy and accelerates convergence.

3. Key Results

The authors successfully performed large-scale calculations on square lattices with periodic boundary conditions ($U/t = 4$) up to $12 \times 12$ sites (144 electrons), a four-fold increase over previous TC-DMRG studies.

• Accuracy Improvements:
  • For equivalent computational effort (comparing TC-DMRG at bond dimension $m$ vs. standard DMRG at $8m$), the TC method reduced the ground-state energy error by a factor of 2.4 to 14.
  • Best Performance: The largest improvement (14x) was observed in a dilute, closed-shell system ($8\times8$, 26 electrons), where the relative error reached 0.02%.
  • Half-Filling: Even for the challenging half-filled case (open-shell), the TC method improved accuracy by 2.4x to 7.0x compared to standard DMRG.
• System Size Scaling:
  • Calculations were performed on systems up to $12\times12$. While the results did not yet match the precision of Auxiliary Field Quantum Monte Carlo (AFQMC) for half-filled systems (due to the difficulty of reducing entanglement within the Fermi surface shell), they significantly outperformed standard DMRG.
  • The method demonstrated that dilute systems are particularly well-suited for TC-DMRG, as the correlator effectively suppresses entanglement away from the Fermi surface.
• Non-Variational Safety: By optimizing $J$ via the residual or variance, the authors ensured that no calculated energy was lower than the AFQMC reference energy, a common pitfall in non-variational TC methods.

4. Significance and Conclusion

• Scalability: The paper demonstrates that the TC method can be scaled to significantly larger systems than previously thought possible by combining optimized MPO construction with entanglement-aware mappings.
• Efficiency: The hybrid MPO algorithm and sparse tensor handling allow for the treatment of Hamiltonians with tens of billions of terms, making high-accuracy calculations on $12\times12$ lattices feasible on standard high-performance computing resources.
• Methodological Advancement: The self-consistent optimization of the correlator parameter $J$ provides a robust framework to mitigate the non-variational nature of the TC method, making it a reliable tool for electronic structure calculations.
• Future Outlook: While the method excels for dilute and closed-shell systems, the authors note that reducing entanglement within the Fermi surface shell at half-filling remains a fundamental challenge. They suggest that future improvements could come from non-uniform bond dimensions, GPU parallelism, or more sophisticated correlators designed specifically for Fermi surface correlations.

In summary, this work establishes Transcorrelated DMRG as a powerful, scalable alternative to standard DMRG and Quantum Monte Carlo for strongly correlated systems, offering a path toward high-accuracy solutions for larger 2D lattices.