[Paper Review] Marshall's sign rule and DMRG acceleration

Ulrich Schollwoeck|arXiv (Cornell University)|Apr 21, 1998

Physics of Superconductivity and Magnetism4 citations

TL;DR

This paper proposes using Marshall's sign rule to dramatically accelerate the Density Matrix Renormalization Group (DMRG) algorithm in antiferromagnetic spin systems. By constructing optimal initial guess vectors for iterative diagonalization based on the sign structure of the ground state, the method reduces the number of required diagonalization steps to as few as 2, accelerating DMRG calculations by up to an order of magnitude, particularly during the growth of long spin chains.

ABSTRACT

In applications of White's Density Matrix Renormalization Group (DMRG) algorithm, computation time is dominated by the diagonalisation of large sparse Hamiltonians by iterative diagonalisation algorithms, whose convergence can be decisively accelerated by the usage of good start vectors. In this paper I show how, using the Marshall sign rule, in a wide class of antiferromagnetic models the number of diagonalization iterations can be reduced below 10, sometimes down to 2, accelerating the DMRG by an order of magnitude. This acceleration, applicable during the growth of long chains, complements the acceleration procedure proposed by White. To illustrate the feasibility of the approach, I show how it performs if applied to the calculation of the Haldane gap for S=2.

Motivation & Objective

• To reduce the computational cost of iterative diagonalization in DMRG, which dominates runtime in large-scale simulations.
• To improve convergence speed of the DMRG algorithm by providing high-quality initial vectors for the diagonalization process.
• To apply Marshall's sign rule—a known property of ground states in antiferromagnetic Heisenberg models—to construct physically motivated starting vectors.
• To demonstrate the feasibility and impact of this acceleration in a non-trivial quantum spin system, specifically the S=2 Haldane chain.

Proposed method

• Leverage Marshall's sign rule, which dictates the sign pattern of the ground state wavefunction in antiferromagnetic Heisenberg models, to construct a physically informed initial vector for iterative diagonalization.
• Use the sign structure from the Marshall rule to build a starting vector that closely approximates the true ground state, reducing the number of Krylov subspace iterations needed.
• Integrate this improved initial vector into the standard DMRG algorithm during the chain growth phase, where diagonalization is most time-consuming.
• Apply the method to the S=2 Haldane chain to test its performance in a system with a known Haldane gap.
• Compare convergence rates (number of iterations) between standard DMRG and the modified version using the sign-rule-based initial vector.

Experimental results

Research questions

• RQ1 Can Marshall's sign rule be used to generate effective initial vectors that accelerate iterative diagonalization in DMRG?
• RQ2 To what extent can the number of diagonalization iterations be reduced using sign-structured initial vectors?
• RQ3 Does this acceleration method remain effective and efficient during the iterative growth of long quantum spin chains?
• RQ4 How does the performance of DMRG with sign-rule initial vectors compare to standard DMRG in a physically relevant model like the S=2 Haldane chain?

Key findings

• The use of Marshall's sign rule to construct initial vectors reduces the number of iterative diagonalization steps to fewer than 10, and in some cases as low as 2, significantly accelerating the DMRG process.
• The method achieves an order-of-magnitude speedup in DMRG calculations, particularly during the chain growth phase where diagonalization dominates runtime.
• The acceleration is robust and effective in the S=2 Haldane chain, a system with a non-trivial Haldane gap, demonstrating practical applicability.
• The approach complements White's original DMRG acceleration techniques by improving the convergence of the diagonalization step, not the algorithmic structure itself.