Let the Hamiltonian Speak for Itself

The challenge

Strongly correlated electron systems are a grand challenge of condensed matter physics: the models describing high-temperature superconductivity and heavy-fermion behavior have no general exact solution. The Dynamical Cluster Approximation solves them numerically by embedding a small cluster in a self-consistent mean-field bath, with Quantum Monte Carlo sampling at its core. DCA++ is its high-performance implementation, developed at Oak Ridge National Lab.

The incredible precision of DCA++ comes at a cost: a typical production run consumes tens of thousands of compute hours for a single calculation.

30,000

GPU-hours per typical run

The object at the core of the Dynamical Cluster Approximation is a cluster of electron sites in momentum space. On each cluster site, the Monte Carlo solver produces an independent, noisy estimate of the Green’s function G.

Figure 1

The Brillouin zone, divided into 16 cluster momenta. Hollow points on the −π edges are periodic images of filled ones. Dashed lines mark select mirror and rotation axes of the D4 point group.

Symmetry buys free samples

Those sixteen estimates are not independent of one another. The point group of the lattice maps cluster momenta onto each other, and symmetry-equivalent points must carry the same value of G. Averaging an orbit of m points is therefore worth up to m independent measurements.

Figure 2

−0.38 − i0.58−0.44 − i0.62

−0.41 − i0.60

one orbit, m = 8 — eight noisy samples, one value

One orbit of eight symmetry-equivalent momenta. Each carries its own noisy estimate; symmetry says they are all measurements of one number.

The variance of G falls accordingly, where N raw samples gives:

Var[Ĝ]=σ²_rawNsymmetrize →σ²_rawmN

“Up to” turns out to be an important qualifier. The variance reduction is capped by m, but it is realized only when the noise on orbit-mates is perfectly uncorrelated. As we will see below, Monte Carlo noise is correlated to different degrees on different models.

The algorithm

The task for DCA++, then, is to automatically identify which symmetries apply to the material model being studied, and impose those symmetries on the Green’s function estimates it produces.

Enumerate symmetry candidates from the theory-limited pool — D₄ ∪ D₆.
Test each against the lattice vectors and the cluster definition.
Solve for the signed permutation matrix U_S on the orbital basis.

An operation survives only if it leaves the Hamiltonian invariant — not just the geometry, but the full matrix structure on the orbital basis:

H(sk)=U_SH(k)U_S^†

Multi-orbital models are the real payoff — In a simple single-band model (one electron per cluster site), geometry alone can identify the symmetry. In a multi-band model (multiple electrons per site, expressing richer physics), an operation may permute orbitals as well as momenta. Those band-permuting operations cannot be validated by the geometry of the lattice - we must consult the Hamiltonian.

Results

Symmetries create a path to an order of magnitude improvement in compute efficiency. We define R, the variance of raw G to the variance of G with symmetries applied - an effective sample multiplier. An R of 2 equates to twice the samples “for free” compared to unsymmetrized G, which we can also think of as a 2x increase in compute efficiency.

R=Var[Ĝ]Var[SĜ]

effective sample multiplier

Figure 3

Square 4×4

1.26

Square 8×8

1.50

Three-band 4×4

3.24

Three-band 6×6

4.49

Kagome 6×6

≈ 6.8

Cubic 6×6×6

≈ 11

1×4×8×12×

MeasuredEstimated ceiling

Effective sample multiplier by model and cluster size. Bars are measured from 1×, where symmetrization buys nothing. Kagome and cubic entries are ceilings implied by their orbit structure, not measured runs.

The single-band square lattice gains little: its orbit-mates are strongly correlated, so eight-fold symmetry buys well under eight-fold variance reduction. The multi-band models are where the method earns its keep — the three-band 6×6 cluster reaches 4.49×, and the orbit structure of the kagome and cubic clusters implies ceilings near 7× and 11×.

Figure pending data

The per-rank scatter needs the ~600 raw and symmetrized estimates from the run behind the poster. Drop them into perRankScatter in src/data/dca-figures.ts and this plot renders itself.

RawSymmetrized

Per-rank estimates of G, three-band 4×4. The raw cloud collapses to a tight cluster once the derived group is imposed — same center, far less spread.

FeAs 2-bandSquare 4×4

R rises as the system cools. The multi-band gain grows with β; the single-band square cluster stays near 1× throughout.

This method helps the most where the physics is most interesting. The multiplier climbs with inverse temperature β, and low temperature is precisely the regime where correlated behavior emerges and where the sampling is most expensive. The derivation costs nothing at runtime — it happens once, at setup, from a matrix the code already has.