[dmftproj] eliminate single-precision loss in the projector construction by krystophny · Pull Request #14 · TRIQS/dftkit

krystophny · 2026-06-16T15:06:12Z

Fixes double-precision values being rounded by default-kind CMPLX constructors in native dmftproj. Angular transforms, spin phases, tetrahedron weights and Loewdin eigenvalues now retain double precision. The shipped p/f templates also retain full-precision coefficients, and the basis reader accepts their longer records.

Adds a native-producer regression using the retained full-rank CaOs2 inputs and a full-precision external d template. It checks symmetry unitarity at 1e-12 independently of golden files. The test fails with the pre-fix producer and passes with this fix in local CMake/CTest. Full TRIQS tests run in CI.

Wien2k supplies the d template externally: this patch cannot recover digits already truncated in an installed template. #12 compares the Python generators against the fixed producer with full-precision templates. The old narrow-window CaOs2 accuracy explanation is superseded by the current 50-band, full-rank fixture; rank-deficient Loewdin output is not a meaningful precision oracle.

Chris&AI

…sion fix) Use the exact analytic cubic harmonics (reptrans no longer applies the legacy float32 cast) and call the same LAPACK routine the Fortran does for the Loewdin inverse (scipy ZHEEV('V','U'), not numpy's ZHEEVD). Regenerate the references from the precision-fixed dmftproj (PR TRIQS#14) with full-precision templates. Against that reference the port is now exact: symqmc 1.5e-14, sympar 1.9e-14, parproj 1.5e-14 (machine precision), oubwin byte-identical. ctqmcout stays at ~6.6e-7: its overlap is rank-deficient (20 correlated spin-orbitals, 17 bands), so O^{-1/2} amplifies the last-ULP libm difference between gfortran and numpy by ~1e8; dmftproj reproduces itself across BLAS but no cross-language port can match it below that floor without changing the (preserved) algorithm. Tolerances tightened to 1e-11 accordingly; ctqmcout documented at 1e-6. Part of TRIQS#7.

…sion fix) Use the exact analytic cubic harmonics (no float32 cast) and call the same LAPACK/BLAS routines the Fortran does for the Loewdin step (scipy ZHEEV('V','U') and ZGEMM with the matching trans flags, not numpy's ZHEEVD / conj-transpose copies). Regenerate references from the precision-fixed dmftproj (PR TRIQS#14) with full-precision templates. The port is now exact to machine precision on every output: oubwin byte-identical symqmc 1.5e-14 sympar 1.9e-14 parproj 1.5e-14 ctqmcout 1.5e-14 (full-rank window: 50 bands > 20 correlated spin-orbitals, so the Loewdin overlap O is non-singular and its eigenvectors are unique. The physical narrow-window CaOs2 makes O rank-deficient; O^{-1/2} then amplifies the last-ULP libm difference, a property of that degenerate case, not the port.) Tolerances tightened to 1e-11. Part of TRIQS#7.

…fixtures The committed CaOs2 dmftproj outputs are now full double precision (precision fix TRIQS#14 + exact Python TRIQS#12), so the converter golden h5 is regenerated to match.

run_dmftproj writes every case.* file the converter reads (band mode writes case.outband), a drop-in for the dmftproj Fortran binary. Delete fortran/dmftproj and its build. End-to-end: the driver feeds the converter and the resulting HDF5 matches the reference (h5diff). The Fortran precision fix lives separately in TRIQS#14; this stack reproduces dmftproj to machine precision. Closes the port. Part of TRIQS#7.

run_dmftproj writes every case.* file the converter reads (band mode writes case.outband), a drop-in for the dmftproj Fortran binary. Delete fortran/dmftproj and its build. End-to-end: the driver feeds the converter and the resulting HDF5 matches the reference (h5diff). The Fortran precision fix lives separately in TRIQS#14. Closes the port. Part of TRIQS#7.

run_dmftproj writes every case.* file the converter reads (band mode writes case.outband), a drop-in for the dmftproj Fortran binary. Delete fortran/dmftproj and its build. End-to-end: driver -> converter -> h5diff. The Fortran precision fix is TRIQS#14; non-SO support is TRIQS#15. Closes the port. Part of TRIQS#7.

Generate correlated-shell spinor symmetry matrices from case.dmftsym, case.indmftpr and case.struct using the dmftproj Wigner-D, angular-basis, time-reversal and spin-phase construction. Cover the non-mixing cubic, spin-mixing fromfile (|j,m_j>) and l=0 paths on the CaOs2 fixture. The cubic matrices differ from the shipped native fixture by 8.6e-8 in complex magnitude (6.9e-8 per real/imaginary component). set_ang_trans.f reads the 8-digit cubic template through CMPLX without KIND=8; magnetic operations apply this rounded basis twice. outputqmc.f also evaluates spin phases in single precision. Python uses exact cubic d coefficients and double-precision phases. Reproducing the native rounding in a local check reduces the complex residual to 2.2e-14; the native precision fixes remain separate in TRIQS#14. Generate into a temporary directory and compare the complete symqmc header exactly and the complex matrices at atol=1e-7, rtol=0. Check Python cubic unitarity at 1e-12. This preserves the SOC converter's shared fixture and avoids h5diff's hard-coded array tolerance and the unrelated parproj archive groups. The fromfile and l=0 tests retain their native matrix references. Read generator inputs from the source fixtures instead of copying unused files into the build directory.

…#5) Generate correlated-shell spinor symmetry matrices from case.dmftsym, case.indmftpr and case.struct using the dmftproj Wigner-D, angular-basis, time-reversal and spin-phase construction. Cover the non-mixing cubic, spin-mixing fromfile (|j,m_j>) and l=0 paths on the CaOs2 fixture. The cubic matrices differ from the shipped native fixture by 8.6e-8 in complex magnitude (6.9e-8 per real/imaginary component). set_ang_trans.f reads the 8-digit cubic template through CMPLX without KIND=8; magnetic operations apply this rounded basis twice. outputqmc.f also evaluates spin phases in single precision. Python uses exact cubic d coefficients and double-precision phases. Reproducing the native rounding in a local check reduces the complex residual to 2.2e-14; the native precision fixes remain separate in #14. Generate into a temporary directory and compare the complete symqmc header exactly and the complex matrices at atol=1e-7, rtol=0. Check Python cubic unitarity at 1e-12. This preserves the SOC converter's shared fixture and avoids h5diff's hard-coded array tolerance and the unrelated parproj archive groups. The fromfile and l=0 tests retain their native matrix references. Read generator inputs from the source fixtures instead of copying unused files into the build directory.

krystophny · 2026-10-12T14:29:12Z

GPT Sol rebased this onto current unstable, reviewed the constructor/template changes, and added an independent native unitarity regression. It fails before the fix and passes afterward in local CMake/CTest. Updated the accuracy explanation to distinguish external template truncation and the obsolete rank-deficient fixture. CI is running.

Chris&AI

krystophny · 2026-10-12T17:40:50Z

GPT Sol re-reviewed the native precision changes against the linked context; no further source change was needed. Reused the precision-fixed native executable as the independent oracle for the new band-mode checks (maximum residual 1.2e-16). All four checks on this unchanged PR head are green.

Chris&AI

the-hampel · 2026-10-27T09:32:46Z

Checked locally on top of current unstable (rebases cleanly): clang++ 22.1 / gfortran 15.2 build without warnings, ctest 28/28 including wien2k_native_precision. All CMPLX changes are pure kind promotions; the set_ang_trans.f buffer widening is benign. The native outputs of this build also reproduce the new full-precision references of #12 to ≤2e-15. Approved, we'll squash-merge.

Attribution: this comment was produced by Claude (Opus 5.5) running in Claude Code, driven by @the-hampel.

This was referenced Jun 16, 2026

[wien2k] pure-Python dmftproj driver; remove the Fortran executable #13

Open

[dmftproj] build fromfile basis representation in double precision #6

Closed

Port dmftproj from Fortran to Python (tracking) #7

Open

Preserve double precision in native dmftproj and test symmetry unitarity

9e7ad09

krystophny force-pushed the fix/dmftproj-precision branch from b8515f4 to 9e7ad09 Compare October 12, 2026 14:29

the-hampel merged commit b08c4e5 into TRIQS:unstable Oct 27, 2026
4 checks passed

the-hampel mentioned this pull request Oct 27, 2026

[wien2k] port and verify the non-spin-orbit (ifSO=0) path #15

Merged

Navigation Menu

Repository navigation

[dmftproj] eliminate single-precision loss in the projector construction - #14

[dmftproj] eliminate single-precision loss in the projector construction#14
the-hampel merged 1 commit into
TRIQS:unstablefrom
krystophny:fix/dmftproj-precision

krystophny commented Jun 16, 2026 •

edited

Loading

Uh oh!

krystophny commented Oct 12, 2026

Uh oh!

krystophny commented Oct 12, 2026

Uh oh!

the-hampel commented Oct 27, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

krystophny commented Jun 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

krystophny commented Oct 12, 2026

Uh oh!

krystophny commented Oct 12, 2026

Uh oh!

the-hampel commented Oct 27, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

krystophny commented Jun 16, 2026 •

edited

Loading