Portable numerical support

Both Fortran implementations compile mymats.f90 and random.f90 from this folder. They require MPI and a BLAS/LAPACK implementation. No NAG, EISPACK, LINPACK, or other third-party library source is bundled here.

The original solver used MyMats from Lib_90_new/Modules/mat_mod.f90 for a small set of matrix operations and the ranf interface from Lib_90_new/Ran/ran_imada.f. These two files are newly written implementations of the required interfaces, under the software distribution's MIT license.

Interface	Bundled implementation	Historical backend
mmult(C,A,B)	BLAS ZGEMM	BLAS ZGEMM
diag(A,U,W)	LAPACK ZHEEV, complex Hermitian	EISPACK CH
inv(A,Ainv,det)	LAPACK ZGETRF, ZGETRI	LINPACK ZGEFA, ZGEDI
udv(A,U,D,V,ncon)	LAPACK QR, A = U diag(D) V	NAG QR
ranf(seed)	Explicit 64-bit modular arithmetic	Implicit 32-bit overflow

The uniform random recurrence remains

seed_next = (48828125 * seed) mod 2147483648
u = seed_next / 2147483648

The archived initial seeds and recurrence are retained verbatim. Do not use a zero seed for new runs. Changes in eigensolver roundoff and compiler or BLAS behavior can still change an interacting Monte Carlo trajectory. Recomputed error bars and means should be assessed statistically.

The active stabilized propagation and local-update Fortran sources are unchanged. The udv adapter supports the retained alternative stabilization routine; the production path uses its own pivoted LAPACK QR code.

Build and check

With GNU Fortran, an MPI development package, and BLAS/LAPACK installed:

make -C src/common check
make -C src/number_conserving build
make -C src/pairing build

compiler.mk selects mpifort when available. Compiler and linker settings can be overridden, for example FC=mpifort LDLIBS='-lopenblas'. The GNU build uses -fallow-argument-mismatch for the inherited mpif.h calling convention. For Intel, first load the Intel oneAPI compiler/MPI environment, then use FC=mpiifx; the default library switch is -qmkl.

make check verifies Hermitian eigendecomposition, nonsymmetric complex inversion and determinant, QR reconstruction and orthogonality, and the uniform random sequence. Each mode also retains live physical benchmark checks and observable tests.

Numerical validation

The numerical interface checks passed with Intel mpiifx 2025.2.1 and MKL. Both the number-conserving and finite-pairing live smoke benchmarks passed. The complete number-conserving regression suite also passed: one analytic free-boson case and four interacting cases with 100000 bins each. The number-conserving eight-bin 3x3 free-boson pipeline reproduced the density, energy density, and two plotted structure factors to within 3e-15 of their analytic values. GNU/BLAS build settings are supplied, but a GNU compiler was not available in this validation environment.