Rank-reduced equation-of-motion coupled cluster formalism with full inclusion of triple excitations

This paper introduces a rank-reduced equation-of-motion coupled cluster formalism that utilizes Tucker decomposition to include full triple excitations with $N^6$ computational cost and $N^4$ storage requirements, while maintaining high accuracy comparable to the canonical method across diverse molecular systems.

The Gist

Imagine you are trying to predict the future behavior of a complex machine, like a car engine, by simulating every single atom inside it. In the world of chemistry, scientists use a powerful mathematical tool called Coupled Cluster theory to do exactly this: simulate how electrons move around atoms to understand how molecules behave, especially when they get excited (like when they absorb light).

The most accurate version of this tool, called EOM-CCSDT, is like trying to simulate every single gear, bolt, and spark in that engine simultaneously. It gives incredibly precise results, but it is so computationally heavy that it's like trying to run a supercomputer simulation on a toaster. It only works for tiny molecules because the time and memory required explode as the molecule gets bigger.

Here is what this paper does, explained through simple analogies:

1. The Problem: The "Too Big to Fit" Puzzle

The authors are dealing with a specific part of the simulation called triple excitations. Think of this as the part of the simulation where three electrons move at the same time. In the standard, "perfect" method, the data required to track these three moving electrons grows so fast (like a snowball rolling down a hill) that it becomes impossible to store on a computer for anything larger than a small molecule.

2. The Solution: The "Smart Compression" Trick

The authors invented a new way to handle this data called Rank-Reduced EOM-CCSDT.

Imagine you have a massive, high-resolution photograph of a crowd of people. If you try to print every single pixel, it takes up a huge amount of paper and ink. However, if you look closely, you realize that many pixels are just variations of the same colors and shapes. You can compress the photo by keeping only the most important patterns and describing the rest as "variations of these patterns."

The authors used a mathematical technique called Tucker decomposition to do exactly this with the electron data. Instead of storing every single possible movement of three electrons, they:

• Found the most important "patterns" of movement.
• Stored only those patterns.
• Reconstructed the full picture using those patterns whenever they needed to do a calculation.

3. The Result: A Faster, Smaller Engine

By using this compression trick, the authors achieved two major things:

• Speed: They reduced the time it takes to run the simulation from something that grows exponentially (like $N^8$) to something much more manageable (like $N^6$). This is the difference between waiting a year for a result and waiting a few days.
• Memory: They drastically reduced the amount of computer memory needed, allowing them to simulate larger molecules that were previously impossible to study with this level of accuracy.

4. Is it Accurate? (The "Good Enough" Test)

You might worry that compressing the data loses accuracy. The authors tested this by comparing their "compressed" method against the "perfect" (but too slow) method on a variety of molecules.

• The Analogy: Imagine you are trying to measure the height of a mountain. The "perfect" method measures every inch. The "compressed" method measures the major peaks and valleys and estimates the rest.
• The Finding: The authors found that their compressed method is incredibly accurate. The error introduced by the compression is much smaller than the natural error already present in the standard, non-compressed version of the theory. In other words, the "compression" doesn't ruin the picture; it's just a slightly blurry version of a picture that was already slightly blurry to begin with.
• The Recommendation: They found that by adjusting one simple "knob" (the size of the compressed subspace), they could get results that are almost indistinguishable from the perfect method for most practical purposes.

5. Real-World Tests

To prove their method works, they didn't just look at theory; they ran actual simulations on:

• Magnesium Dimer: They mapped out the energy curves for a magnesium molecule, showing they could predict how it vibrates and holds together, matching experimental data well.
• Ammonia and Fluorine: They simulated a "charge-transfer" event (where an electron jumps from one molecule to another over a distance). This is notoriously difficult for other methods, but their compressed method handled it smoothly, producing clean, continuous curves without glitches.

Summary

In short, this paper presents a smart shortcut. It takes a method that is too expensive to use for big molecules and compresses the data so it becomes affordable, without sacrificing the high accuracy that scientists need. It's like taking a super-detailed 8K movie and compressing it into a high-quality 4K file that still looks amazing but fits on a standard hard drive. This allows chemists to study larger, more complex systems with a level of precision that was previously out of reach.

Technical Summary

Technical Summary: Rank-Reduced Equation-of-Motion Coupled Cluster Formalism with Full Inclusion of Triple Excitations

Problem Statement
The accurate description of molecular excited states, particularly those involving significant double or higher-order excitations, or charge-transfer character, remains a challenge for standard quantum chemical methods. While Time-Dependent Density Functional Theory (TDDFT) is computationally efficient, it lacks systematic improvability and often struggles with charge-transfer states. Coupled Cluster (CC) theory, specifically the Equation-of-Motion (EOM-CC) formalism, offers a systematically improvable path to the exact solution of the electronic Schrödinger equation. However, the inclusion of triple excitations (EOM-CCSDT), which is often necessary for benchmark-quality results or to correct large errors in EOM-CCSD (e.g., in charge-transfer or double-excitation dominated states), scales as $N^8$ with system size. This prohibitive cost restricts EOM-CCSDT to very small systems (typically fewer than a dozen non-hydrogen atoms). Approximate methods like CC3 or CCSD(T) reduce scaling but may not capture the full physics of triple excitations required for certain challenging states.

Methodology
The authors propose a rank-reduced variant of the EOM-CCSDT method (RR-EOM-CCSDT) that reduces the computational scaling to $N^6$ and storage requirements to $N^4$. The core of the formalism relies on the Tucker decomposition of the triply-excited amplitudes tensors for both the ground state ($T_{ijk}^{abc}$) and excited states ($R_{ijk}^{abc}$).

1. Tensor Decomposition: The full-rank amplitudes are approximated as:
   $$T_{ijk}^{abc} \approx t_{xyz} U_{ia}^x U_{jb}^y U_{kc}^z$$
   $$R_{ijk}^{abc} \approx r_{XYZ} V_{ia}^X V_{jb}^Y V_{kc}^Z$$
   Here, $t_{xyz}$ and $r_{XYZ}$ are compressed amplitudes, while $U$ and $V$ are basis tensors spanning the triple-excitation subspaces. The dimensions of these subspaces ($N_{svd}$ and $N_{SVD}$) are controlled parameters, typically scaling linearly with the system size ($N$), rather than the cubic scaling of the full tensor.

2. Subspace Generation (Guessing): To determine the subspaces $U$ and $V$, the authors employ the Higher-Order Orthogonal Iteration (HOOI) procedure. This requires an initial guess for the triple amplitudes. Two guessing strategies are derived via perturbation theory expansion of the EOM-CCSDT equations:

   • Basic Guess: Includes only first-order perturbative terms.
   • Extended Guess: Includes first-order terms plus an approximate second-order contribution (omitting the most expensive term involving the first-order triple amplitudes to maintain efficiency).
     The HOOI procedure iteratively refines these guesses to minimize the least-squares error, yielding the optimal subspaces without ever explicitly constructing the full-rank $N^8$ tensors.

3. Projected Equations: The EOM-CCSDT residual equations are projected onto the compressed subspaces. This transforms the eigenvalue problem into one involving only the compressed amplitudes ($r_{XYZ}$) and a projected residual tensor ($\Omega_{XYZ}$). The derivation ensures that the most expensive terms, which originally scale as $O^3V^5$ (where $O$ is occupied and $V$ is virtual orbitals), are factorized to scale as $O^2V^4$ or $O^3V^3$ depending on the specific term, leading to an overall $N^6$ scaling.

Key Contributions

• Formalism Development: The paper extends previous rank-reduced work (specifically RR-EOM-CC3) to the full EOM-CCSDT level, providing explicit working formulas for the projected triple-amplitude equations and the construction of the compressed residual tensor.
• Scaling Reduction: The method achieves $N^6$ computational scaling and $N^4$ storage scaling, making EOM-CCSDT applicable to significantly larger systems than the canonical $N^8$ approach.
• State-Specific Approach: The method is inherently state-specific, as the triple-excitation subspace ($V$) is optimized for a specific excited state.
• Implementation of Guesses: The authors derive and compare "basic" and "extended" perturbative guesses for the HOOI procedure, analyzing their trade-offs between accuracy and computational cost.

Results
The authors validated the method through three distinct applications:

1. Benchmark Calculations: Tests were performed on ten molecules (e.g., acrolein, benzene, water, nitroxyl) with excited states ranging from single-excitation dominated to double-excitation dominated.

   • Accuracy: Using a subspace size of $N_{SVD} = 2.0 \times N_{MO}$ (where $N_{MO}$ is the number of active molecular orbitals), the mean absolute error (MAE) relative to canonical EOM-CCSDT was found to be approximately 0.008–0.009 eV.
   • Comparison: These errors are several times smaller than the inherent error of the parent EOM-CCSDT method relative to Full Configuration Interaction (FCI) for the same states. The method successfully captures triple-excitation effects even for states where EOM-CCSD fails by several eV.
   • Guess Performance: While the extended guess showed better accuracy for some specific cases (e.g., nitroxyl), it was found to be non-systematic (performing worse for others like glyoxal) and significantly more expensive. The authors recommend the basic guess with a larger subspace size as the more robust and economical default.

2. Magnesium Dimer (Mg$_2$): Potential energy curves (PECs) and spectroscopic parameters ($D_e$, $R_e$, $\omega_e$) were calculated for the first four singlet excited states.

   • The method produced smooth, continuous PECs without unphysical discontinuities, despite the subspaces being re-optimized at each internuclear distance.
   • Results extrapolated to the Complete Basis Set (CBS) limit showed good agreement with experimental data and high-level theoretical benchmarks, particularly for the $A^1\Sigma_u^+$ and $(1)^1\Pi_u$ states.

3. Charge-Transfer Excitation (NH$_3$-F$_2$): The method was tested on a long-range charge-transfer excitation as a function of intermolecular separation (6–100 bohr).

   • The RR-EOM-CCSDT curves remained smooth and physically consistent across the entire range, correctly reproducing the $1/R$ behavior of the charge-transfer state.
   • Errors relative to the canonical limit were generally small (e.g., ~0.03 eV for $N_{SVD}=3N_{MO}$), demonstrating the method's stability for long-range interactions where orbital relaxation is critical.

Significance and Claims
The authors claim that the RR-EOM-CCSDT method provides a practical route to obtaining excitation energies of near-canonical EOM-CCSDT quality for systems where the full method is computationally prohibitive. The method is presented as a significant step forward for:

• Large Systems: Enabling high-accuracy calculations for molecules with more than a dozen non-hydrogen atoms.
• Composite Schemes: Serving as a cost-effective component in composite schemes that account for higher-order correlation effects.
• Future Extensions: Laying the necessary theoretical foundation for rank-reduced methods including quadruple excitations (EOM-CC4), which are expected to deliver near-FCI quality for a broader range of excited states.

The paper emphasizes that while the rank-reduced formalism introduces an approximation, the error introduced is controllable and, with default settings ($N_{SVD} = 2N_{MO}$), is negligible compared to the intrinsic limitations of the parent EOM-CCSDT theory itself.