Skip to content

Benchmark reproduction

The default python3 reproduce.py reruns all 22 BAFQMC/reference calculations, processes the new measurements, and draws the paper benchmark figures. See RESOURCES.md for the 12–24 hour production budget, memory, scratch storage, and continuation commands. The optional --mode plot redraws the small stored dataset. Both paths draw the combined relative-density and the separate attractive total-density benchmark.

Data and reference conventions

Figure / row Varying parameter Fixed parameters Reference
Main Fig. 2(a–d) U1 = 0, 0.25, 0.5, 0.75, 1, 1.25, 1.5, 2 U2 = 0, mu = -3.5, beta = 4 Particle-shell ED; analytic free-boson reference at U1 = 0
Main Fig. 2(e–h) Delta = 0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3 U1 = 1, U2 = 0, mu = -5, beta = 4 Dense ED with nmax = 3 and ncut = 4
Total-density figure (a–d) U2 = -0.6, -0.5, -0.4, -0.3, -0.2, -0.1, 0 U1 = 1, mu = -7, beta = 1 Low-density finite-occupation-window reference for U2 < 0

All inputs, ED records, processed tables, and plot metadata use the manuscript notation directly: U1 is the repulsive relative-density coefficient and U2 is the attractive total-density coefficient.

--scope combined selects the 15 relative-density and pairing points, and --scope total_density selects the seven total-density points. The default --scope all includes both. Scope and solver selection are independent:

Selection Points Scan
--scope combined 15 U1 and Delta
--scope combined --model number_conserving 8 U1
--scope combined --model pairing 7 Delta
--scope total_density 7 U2
--model number_conserving 15 U1 and U2

The U2=0 benchmark points satisfy mu < -3*t - abs(Delta) at every plotted point. As proved in the study's convergence subsection "Convergence throughout the auxiliary-field domain", its quadratic propagation and unitary relative-density HS factors give a finite trace throughout the auxiliary-field domain. This result applies to the U2=0 scan; the U2<0 scan uses the finite-occupation references encoded in its data records.

The pairing implementation writes the real pair term with a positive input coefficient, whereas the manuscript writes it as -Delta (b^+ c^+ + b c). The phase convention c_code = -c_paper relates these expressions at the same positive scan value. Density, total physical energy, and the two plotted structure factors are invariant under this transformation. The solver's anomalous pair amplitude pair_equal changes sign when expressed in manuscript operators; see the pairing conventions.

All points have t = 1, a periodic 3 x 3 triangular lattice, Delta tau = 0.01, 100000 measurement bins, and one MPI rank. The error bars are standard errors from ten consecutive blocks of 10000 bins. Warmup and stabilization settings, proposal widths, and exact initial seeds are in each inputs/<case>/ directory. The paired seeds.txt contains a list of rank seeds; the paper uses its first entry. It is not a multiword state for one random-number stream.

The four measurement files are density_total, energy_density, sf_K, and dw_K. Structure-factor files store real and imaginary components. The plots use their real parts and show the physical total energy as -E = -9 e, with the energy standard error multiplied by 9. The chemical-potential term is excluded from E. No statistical acceptance criterion removes points during reproduction.

The attractive-U2 reference uses the low-density occupation window: six completed particle shells, with a configured maximum of eight. This is the finite-window comparison encoded by low_density_cutoff_accepted in the ED records. The U1 = 0 reference in the relative-density interaction scan is computed analytically; all other reference values are read from the stored ED results.

Layout

data/index.json                  selected cases, parameters, plotted values, and metadata
data/observables.csv             compact processed means, SEM, and ED values
data/block_means.csv             ten block means per case and observable
data/<model>/manifest.json       executable paper campaign
data/<model>/inputs/<case>/      exact initial input files and ED parameters
data/<model>/ed/<case>.json      compact ED reference results
data/<model>/raw/<case>.tar.gz   optional local raw chains, ignored by Git
analysis.py                     raw-data statistics, references, tables, plot dispatch
plot_manuscript.py               relative-density/pairing grid and total-density scan
plot_number.py                  2 x 4 number-conserving layout
plot_pairing.py                 1 x 4 paired layout
plot_style.py                   shared figure style

The tracked data package is approximately 250 kB. Large raw measurement files are generated by BAFQMC and excluded from Git. The small ED JSON files contain reference expectation values and cutoff metadata. Stored block means permit independent reconstruction of the plotted mean and standard error without shipping the full Monte Carlo chains. --mode check checks finite records, ED parameters, and reconstruction of the plotted means and SEM without making figures. --mode raw reconstructs the statistics from optional local raw chains when they are present locally.

The current outputs are figures/benchmark_combined.pdf (two rows) and figures/benchmark_total_density.pdf (one row), with PNG previews alongside them. Only the figures and rows covered by the selected scope and solver are drawn. An individual combined-figure row is written as figures/benchmark_combined_number_conserving.pdf or figures/benchmark_combined_pairing.pdf, preserving panel labels (a–d) or (e–h), respectively. The layout modules remain available as reusable plotting helpers.

Generated tables include benchmark, section, figure, figure_row, and panels to locate each point in the current manuscript. scan_parameter is the input key and plotted symbol. The JSON records collect these layout fields under manuscript. Case IDs follow the same U1/U2 notation; row identifies the benchmark family and figure_row identifies the current figure arrangement.

Fresh calculations

--mode full (the default) builds each solver, executes the paper campaign, runs ED, reblocks the new measurements, and draws both figures from the new results. The tracked processed data stay fixed in data/. Output goes to the directory selected by --output and includes the numerical differences from ED.

Fresh Monte Carlo trajectories can differ with compiler and numerical library: the tracked data remain the exact source for regenerating the paper curves. PDF metadata and font-library versions can also change PDF byte hashes even when the plotted data and layout are identical.

For separate stages, --mode dqmc and --mode ed write into output/runs/. After both stages, --mode analyze reblocks their outputs and draws the figures. Use the same --scope and --model for each stage and for --resume. Each solver's run_paper.py --help also exposes case selection or campaign configuration and an analysis entry point. Copy the supplied inputs into a separate configuration directory when creating a new model campaign. All source files, campaign inputs, processed data, and reference implementations required for reproduction are included here. Production generates fresh raw measurements directly from the included inputs.