A Lanczos-based algorithm for sum-over-states calculations of NMR spin--spin coupling constants at the RPA level of theory: The Fermi-contact term

This study demonstrates that employing a Lanczos algorithm within the RPA framework significantly improves the efficiency of calculating Fermi-contact contributions to NMR spin-spin coupling constants, achieving convergence with less than 50% of the excited states required by traditional Davidson-based sum-over-state methods.

The Gist

The Big Picture: Listening to the Atomic Orchestra

Imagine you are trying to understand a complex piece of music played by an orchestra. In the world of chemistry, this "music" is the Nuclear Magnetic Resonance (NMR) signal. Specifically, scientists are looking at how two atoms in a molecule "talk" to each other through their magnetic spins. This conversation is called a Spin-Spin Coupling Constant (SSCC).

To understand why the atoms talk the way they do, scientists try to break the music down into individual notes. In this case, the "notes" are excited states—temporary, high-energy configurations the electrons in the molecule can jump into.

The Problem:
Traditionally, to get the perfect answer, you have to listen to every single note the orchestra can play, from the lowest bass drum to the highest, squeaky violin. For large molecules, there are millions of these "notes." Calculating all of them is like trying to listen to a 10-hour symphony just to hear the first 30 seconds of the melody. It takes too long and costs too much computing power.

The Old Way (The Davidson Algorithm):
Previously, scientists used a method called the Davidson algorithm. Imagine this as a librarian who only looks for books starting from the bottom shelf (the lowest energy notes) and works their way up.

• The Issue: The librarian keeps finding that the very top shelves (the highest energy notes) actually contain the most important clues for the melody. So, the librarian has to keep climbing until they reach the very top of the library. They can't stop early because they don't know if the last few books on the top shelf will change the story.

The New Solution: The Lanczos Algorithm

The authors of this paper introduced a new method called the Lanczos algorithm.

The Analogy: The Elevator vs. The Stairs

• The Davidson Algorithm (Stairs): You start at the bottom floor and walk up one step at a time. You don't know what's on the top floor until you get there.
• The Lanczos Algorithm (The Elevator): Imagine an elevator that starts in the middle of the building but instantly shoots to the top floor and the bottom floor simultaneously. Then, it comes down one floor from the top and goes up one floor from the bottom.

Why is this better?
The Lanczos algorithm grabs the "extreme" notes (the very highest and very lowest energy states) first.

• In the world of these magnetic conversations, the highest energy notes (which the old method ignored until the very end) are actually the ones that matter most for the "Fermi-contact term" (the loudest part of the signal).
• Because Lanczos grabs these important high-energy notes immediately, it can give you a very accurate answer after listening to only 40% to 50% of the total orchestra. The old method often needed 100%.

What Did They Do?

The researchers took this "elevator" method and applied it to 17 different molecules (ranging from simple things like water and methane to slightly more complex ones like ethane).

1. The Test: They calculated the magnetic "conversation" between atoms using the new method, stopping after 10%, 20%, 30%... all the way to 100% of the notes.
2. The Result: For most molecules, they found that once they included about half of the possible excited states, the answer stopped changing. It had "converged."
   • For simple molecules like ethane, they only needed 20% of the states to get a perfect answer.
   • For a few tricky molecules with heavy atoms (like Phosphorus or Chlorine), they needed about 60%.
3. The "Spikes" Problem: With the old method, if you stopped halfway, the answer would jump around wildly (spikes) because you missed a crucial high-energy note. With the new Lanczos method, once the answer settles, it stays settled. It's stable.

A Crucial Detail: Choosing the Right "Microphone"

There is one catch. To make the elevator work, you have to tell it where to start.

• If you are trying to hear the conversation between Atom A and Atom B, you need to point your "microphone" (the starting vector) at Atom A or Atom B.
• If you point the microphone at a random Atom C, the elevator might take a weird route, and the answer won't be as accurate.
• The Good News: Even though you have to do this calculation separately for each pair of atoms, it's actually a benefit! Because each pair is independent, you can run all these calculations at the same time on different computers (parallelization).

The Bottom Line

This paper proves that we don't need to listen to the entire 10-hour symphony to understand the melody. By using the Lanczos algorithm, we can listen to the most important "extreme" notes first.

• Old Way: Listen to 100% of the notes, slowly, from bottom to top. (Slow, expensive, unstable).
• New Way: Listen to the top and bottom notes first. (Fast, cheap, stable).

This allows scientists to study much larger and more complex molecules than ever before, giving us a deeper understanding of how molecules behave, without needing a supercomputer to do the math.

Technical Summary

1. Problem Statement

The paper addresses a significant computational bottleneck in calculating Nuclear Magnetic Resonance (NMR) spin–spin coupling constants (SSCCs) using the Sum-Over-States (SOS) approach within the Random Phase Approximation (RPA) or Time-Dependent Hartree-Fock (TD-HF) framework.

• The Challenge: To accurately calculate the dominant Fermi-contact (FC) term of SSCCs, traditional SOS methods require summing contributions over all excited states (including high-energy pseudo-states representing the continuum).
• Limitations of Current Methods: Previous studies using the Davidson algorithm (which converges the lowest eigenvalues first) demonstrated that truncating the sum of excited states leads to poor convergence. Even high-lying excited states contribute significantly to the FC term, often causing large fluctuations ("spikes") in the calculated coupling constants until nearly 100% of the states are included. This makes SOS analysis computationally prohibitive for larger molecules.
• The Goal: The authors aim to determine if the Lanczos algorithm, which converges eigenvalues from both the top and bottom of the spectrum simultaneously, can achieve converged FC values using a significantly smaller subset of excited states (pseudo-states).

2. Methodology

The study extends a recent implementation of the Lanczos algorithm (originally developed for mean excitation energies) to calculate SSCCs within the Dalton quantum chemistry program.

• Theoretical Framework:
  • Calculations are performed at the RPA/TD-HF level.
  • The SSCC is decomposed into four terms: Fermi-contact (FC), Spin-dipolar (SD), Paramagnetic Spin-Orbit (PSO), and Diamagnetic Spin-Orbit (DSO). The study focuses primarily on the FC term.
  • The FC term is expressed as a sum over contributions from pairs of occupied and virtual orbitals, requiring excitation energies ($\omega_n$) and eigenvectors ($X_n, Y_n$).
• Algorithmic Implementation:
  • Paired Structure Preservation: The RPA eigenvalue problem involves a matrix with a specific paired structure (positive and negative eigenvalues). The authors modified the Lanczos algorithm to preserve this structure by constructing block Lanczos vectors ($Q_k, \tilde{Q}_k$) that mirror the excitation/de-excitation symmetry of the RPA eigenvectors.
  • Spectrum Convergence: Unlike the Davidson algorithm (which finds the lowest $N$ eigenvalues), the Lanczos algorithm approximates the largest and smallest eigenvalues simultaneously. This allows it to capture high-energy pseudo-states early in the iteration process.
  • Eigenvector Reconstruction: To analyze contributions from specific orbital pairs (CLOPPA analysis), the algorithm reconstructs the full-space eigenvectors from the reduced Lanczos basis vectors.
  • Start Vectors: A critical methodological choice was the selection of property gradient vectors as start vectors. The authors found that using the FC property gradient of one of the coupled nuclei (rather than a generic dipole operator) yields the fastest convergence.

3. Key Contributions

• Algorithm Extension: Successfully adapted the block Lanczos algorithm for RPA eigenvalue problems to calculate SSCCs, specifically handling the reconstruction of full-space eigenvectors required for orbital-pair analysis.
• Convergence Strategy: Demonstrated that the Lanczos algorithm converges the FC term much faster than the Davidson algorithm by prioritizing the inclusion of high-energy states that are crucial for the FC contribution.
• Start Vector Optimization: Identified that using the FC property gradient of the specific coupled atoms as the start vector is essential for rapid convergence, particularly for degenerate or equivalent bonds.

4. Results

The method was tested on 17 molecules containing first, second, and third-row atoms (e.g., $BH_3$, $C_2H_2$, $CH_4$, $H_2O$, $PH_3$, $SiH_4$).

• Convergence Efficiency:
  • For most coupling constants, the FC term converged to within 0.5 Hz of the full-space linear response value using less than 50% of the total excited states.
  • One-bond couplings: Converged at ~41% on average. X-H couplings showed a wider spread (20% to >60%) compared to X-Y couplings.
  • Two-bond and Vicinal couplings: Generally converged around 42–46%.
  • Geminal couplings: Converged at ~44% when using the FC gradient of the coupled hydrogen atoms as the start vector.
  • Exceptions: Molecules containing third-row atoms (e.g., $PH_3$, $HCl$, $H_2S$, $H_2O$) required higher percentages (up to 61%) for convergence.
• Comparison with Davidson Algorithm:
  • Stability: The Lanczos algorithm produces stable results once the threshold is reached. In contrast, the Davidson algorithm exhibits large oscillations and "spikes" even at high iteration counts because it misses high-energy states until the very end.
  • Speed: For ethane ($C_2H_6$), Lanczos converged at ~20% of states, whereas Davidson required ~75% and still showed instability.
• Start Vector Sensitivity: The convergence rate is highly dependent on the start vector. Using the FC gradient of the relevant coupled atom significantly outperforms using gradients from other atoms or generic operators.

5. Significance

• Computational Feasibility: This work enables CLOPPA (Contributions from Localized Orbitals within the Polarization Propagator Approach) analyses for larger molecules that were previously computationally inaccessible due to the cost of calculating all excited states.
• Reliability: It provides a robust method for truncating SOS sums without the risk of the erratic fluctuations seen in Davidson-based truncations.
• Future Applications: The success of the Lanczos approach for the FC term suggests it could be extended to calculate the SD and PSO terms, potentially revolutionizing the efficiency of NMR parameter calculations in quantum chemistry.
• Parallelization: While the need for specific start vectors requires separate calculations for each coupling constant, this allows for trivial parallelization of the workload.

In summary, the paper demonstrates that the Lanczos algorithm is a superior alternative to the Davidson algorithm for SOS calculations of SSCCs, offering faster convergence, greater stability, and the ability to analyze coupling pathways in larger molecular systems.