Method for making 2-electron response reduced density matrices approximately N-representable

Caitlin Lanssens, Paul W. Ayers, Dimitri Van Neck, Stijn De Baerdemacker, Klaas Gunst, Patrick BultinckThe Journal of Chemical Physics 2018 · DOI

TL;DR

This paper introduces a novel algorithm to "fix" approximate 2-electron reduced density matrices (2-RDMs) generated by response theory, making them approximately N-representable. This is crucial because these response-theory 2-RDMs, while computationally convenient, don't correspond to real physical systems. The new method ensures the corrected 2-RDMs satisfy essential mathematical properties, improving the reliability of quantum chemical calculations.

What Problem Does This Paper Address?

In quantum mechanics, accurately describing the behavior of electrons in molecules is paramount. The full picture of an N-electron system is captured by its wave function, but this is often computationally intractable for all but the smallest systems. An alternative is to work with reduced density matrices (RDMs), which are simpler objects that still contain significant information about the system. Specifically, the 2-electron RDM (2-RDM) contains all the information needed to calculate the system's energy.

However, many advanced quantum chemistry methods, like coupled cluster (CC) and geminal-based approaches (e.g., AP1roG), rely on approximations. When direct computation of the 2-RDM is too expensive, researchers often turn to "response theory" to obtain an approximate 2-RDM. The critical problem is that these response-theory 2-RDMs are often not N-representable. This means they don't correspond to any physically realizable N-electron wave function. This lack of N-representability can lead to unphysical results and errors in subsequent calculations. This paper aims to bridge this gap by providing a method to correct these non-N-representable 2-RDMs.

Methods

The core of this paper is a new algorithm designed to make approximate 2-RDMs approximately N-representable. The approach is framed as an optimization problem: find a "target" 2-RDM that is as close as possible (in the Frobenius norm sense) to the initial, non-N-representable response-theory 2-RDM, while satisfying crucial N-representability conditions.

The key N-representability conditions addressed are:

  • Symmetry: The 2-RDM must have the correct symmetry properties.
  • Normalization: The trace of the 2-RDM must correspond to the correct number of electrons.
  • Positivity Conditions (P, Q, G): These are a set of mathematical constraints derived from the requirement that the expectation value of any positive semidefinite operator must be non-negative. These conditions ensure that the 2-RDM can be derived from a valid wave function.

The algorithm iteratively "fixes" these properties. It involves:

  1. Symmetrization: Ensuring the 2-RDM has the correct symmetry.
  2. Eigenvalue Decomposition and Shifting: This is the core of the "fixing" process. Eigenvalues of the 2-RDM (and related P, Q, G matrices) are analyzed. Negative eigenvalues, which violate positivity, are shifted. A constant shift is applied to all eigenvalues to maintain the correct trace. This process is repeated until the eigenvalues are non-negative and the trace is correct.
  3. Specialized Algorithm for Seniority-Zero: The authors also developed a tailored optimization procedure specifically for seniority-zero 2-RDMs, which are relevant for methods like DOCI (doubly occupied configuration interaction). This specialized algorithm leverages the simplified structure of seniority-zero 2-RDMs.

The authors frame this problem as a semidefinite programming problem but opt for a computationally faster "fixing" procedure, inspired by Higham's work on finding the closest positive semidefinite matrix.

Datasets

Not discussed in the paper. The paper focuses on developing and testing a computational method, rather than analyzing specific pre-existing datasets. The "datasets" in this context are the output 2-RDMs from various quantum chemistry calculations (AP1roG, DOCI) for different molecular systems.

Key Insights and Ideas

  • The "N-representability problem" is a fundamental hurdle: Response theory provides computationally convenient 2-RDMs, but they often lack the mathematical properties required for them to represent a real physical system. This paper directly tackles this issue.
  • A "fixing" procedure is more practical than full optimization: While a full semidefinite programming approach is theoretically ideal, the authors demonstrate that an iterative "fixing" algorithm is computationally more feasible and efficient, especially when starting with an already approximate 2-RDM.
  • Eigenvalue manipulation is key: The core of the N-representability fixing lies in carefully shifting and ensuring the positivity of eigenvalues of the 2-RDM and related matrices (P, Q, G).
  • Seniority-zero systems have special properties: The authors highlight that 2-RDMs from seniority-zero wave functions (like DOCI) have a simplified structure, allowing for a more specialized and efficient optimization algorithm.
  • The cost of N-representability: Enforcing N-representability can lead to an increase in energy compared to the initial response-theory result. This is seen as an "energetic cost" for ensuring the physical realizability of the 2-RDM.

Key Results

  • Successful N-representability enforcement: The proposed algorithm successfully makes non-N-representable 2-RDMs approximately N-representable, satisfying symmetry, normalization, and the P-, Q-, and G-conditions.
  • Computational efficiency: The developed algorithm is significantly faster than exact methods for obtaining N-representable 2-RDMs.
  • System-specific performance:
    • For systems where AP1roG is accurate (like the dissociation of H8), the initial response 2-RDMs are already close to N-representable, and the "fixing" procedure introduces only small energy changes.
  • For stretched molecular geometries and systems with strong correlation (like the attractive pairing Hamiltonian), the initial response 2-RDMs deviate more significantly from N-representability, and the fixing procedure leads to larger energy increases.
  • The specialized DOCI 2-RDM algorithm shows higher energy increases for stretched geometries compared to the general 2-RDM algorithm, but it is restricted to seniority-zero space.
  • Comparison with exact solutions: In the case of the pairing Hamiltonian, the authors show that while the fixed 2-RDM becomes N-representable, it doesn't necessarily become closer to the exact 2-RDM, suggesting that N-representability and exactness are related but distinct concepts.

Limitations and Future Work

  • Approximation: The method provides approximately N-representable 2-RDMs, not strictly N-representable ones. The goal is to satisfy the most important conditions.
  • Energetic Cost: Enforcing N-representability can lead to an increase in the calculated energy, which might be significant for systems with strong correlation or stretched geometries.
  • Seniority-Zero vs. General: The specialized DOCI algorithm, while efficient for seniority-zero systems, can lead to larger energy increases than the general algorithm for certain cases.
  • Future Work: The authors suggest that the algorithm could be applied to other quantum chemistry methods that produce non-N-representable 2-RDMs. Further exploration of the trade-off between computational cost, accuracy, and the degree of N-representability is also implied. The paper also hints at the possibility of changing the order of fixing P, Q, and G conditions, suggesting room for further optimization of the algorithm itself.

Summary updated

Abstract

In methods like geminal-based approaches or coupled cluster that are solved using the projected Schrödinger equation, direct computation of the 2-electron reduced density matrix (2-RDM) is impractical and one falls back to a 2-RDM based on response theory. However, the 2-RDMs from response theory are not N-representable. That is, the response 2-RDM does not correspond to an actual physical N-electron wave function. We present a new algorithm for making these non-N-representable 2-RDMs approximately N-representable, i.e., it has the right symmetry and normalization and it fulfills the P-, Q-, and G-conditions. Next to an algorithm which can be applied to any 2-RDM, we have also developed a 2-RDM optimization procedure specifically for seniority-zero 2-RDMs. We aim to find the 2-RDM with the right properties which is the closest (in the sense of the Frobenius norm) to the non-N-representable 2-RDM by minimizing the square norm of the difference between this initial response 2-RDM and the targeted 2-RDM under the constraint that the trace is normalized and the 2-RDM, Q-matrix, and G-matrix are positive semidefinite, i.e., their eigenvalues are non-negative. Our method is suitable for fixing non-N-representable 2-RDMs which are close to being N-representable. Through the N-representability optimization algorithm we add a small correction to the initial 2-RDM such that it fulfills the most important N-representability conditions.
BibTeX
@article{lanssens2018method,
  author = {Caitlin Lanssens and Paul W. Ayers and Dimitri Van Neck and Stijn De Baerdemacker and Klaas Gunst and Patrick Bultinck},
  title = {{Method for making 2-electron response reduced density matrices approximately <i>N</i>-representable}},
  journal = {The Journal of Chemical Physics},
  year = {2018},
  doi = {10.1063/1.4994618},
}
Figure