From 18ec7ff94cfdb751301d00eb5980c3936ee2324b Mon Sep 17 00:00:00 2001 From: Cail Daley Date: Fri, 10 Jul 2026 03:25:37 +0200 Subject: [PATCH 1/3] =?UTF-8?q?feat(twopoint):=20SACC=20=E2=86=92=202pt-FI?= =?UTF-8?q?TS=20converter,=20byte-compatible=20with=20cosmosis=5Ffitting?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Add sp_validation.twopoint_convert.sacc_to_twopoint_fits, converting an analysis SACC into the CosmoSIS 2pt-FITS that 2pt_like (and Sacha's rho/tau 2pt_like_xi_sys fork) reads. Output reproduces today's cosmo_inference cosmosis_fitting.py assembly HDU-for-HDU and byte-for-byte: NZDATA, XI_PLUS/MINUS, CELL_EE, the blocked COVMAT (STRT_i offsets) and separate COVMAT_CELL, and the TAU_{0,2}_PLUS + verbatim RHO_STATS tables. The data vector and covariance are permuted from SACC's pair-major order to the 2pt-FITS type-major layout via s.indices. The tau covariance is laid in as one contiguous [tau_0+; tau_2+] block, preserving the tau_0<->tau_2 cross-correlation that covdat_to_fits carries. Rho/tau HDUs need the rho/tau sidecar FITS: the RHO_STATS varrho_* variance columns are not stored in the analysis SACC, so the converter copies them verbatim (never fabricates them); xi, Cl, n(z) and the covariance are fully reconstructed from SACC alone. Tests: - test_twopoint_convert.py: byte-equal vs current cosmosis_fitting.py builders on deterministic synthetic inputs for all three product shapes (plain xi; xi+Cl; xi+rho/tau), plus the tau_0<->tau_2 cross-correlation and perturbation teeth. - test_twopoint_convert_realdata.py (skipif on candide paths): build a SACC from a real product's own contents, convert, byte-compare vs the current writer. Observed byte-equal for SP_v1.4.6_leak_corr and a glass_mock sibling; the stale on-disk files differ only by an extra CELL_BB HDU (older script version) and ~1e-17 float noise on bin edges, documented and guarded. Co-Authored-By: Claude Fable 5 --- .../tests/test_twopoint_convert.py | 378 ++++++++++++++++++ .../tests/test_twopoint_convert_realdata.py | 268 +++++++++++++ src/sp_validation/twopoint_convert.py | 330 +++++++++++++++ 3 files changed, 976 insertions(+) create mode 100644 src/sp_validation/tests/test_twopoint_convert.py create mode 100644 src/sp_validation/tests/test_twopoint_convert_realdata.py create mode 100644 src/sp_validation/twopoint_convert.py diff --git a/src/sp_validation/tests/test_twopoint_convert.py b/src/sp_validation/tests/test_twopoint_convert.py new file mode 100644 index 00000000..5e599717 --- /dev/null +++ b/src/sp_validation/tests/test_twopoint_convert.py @@ -0,0 +1,378 @@ +"""Byte-compare tests for the SACC -> 2pt-FITS converter. + +The converter (:mod:`sp_validation.twopoint_convert`) must reproduce the CosmoSIS +2pt-FITS that ``cosmo_inference/scripts/cosmosis_fitting.py`` assembles today, +so the inference chain (``2pt_like`` and Sacha Guerrini's rho/tau +``2pt_like_xi_sys`` fork) runs untouched behind it. The strongest possible check +is *byte* equality, and astropy writes FITS deterministically, so that is what we +assert: build a reference with the current script's own HDU-builder functions on +deterministic synthetic inputs, build a SACC from those same inputs via +:mod:`sp_validation.sacc_io`, convert it, and compare the two files byte for byte. + +Three configurations pin the three product shapes today's ``__main__`` emits: + +1. **plain xi** -- PRIMARY, NZ_SOURCE, COVMAT, XI_PLUS, XI_MINUS. +2. **xi + pseudo-Cl** -- adds COVMAT_CELL and CELL_EE (the harmonic block + ``2pt_like`` reads; the script builds CELL_BB and discards it, so the + converter does too). +3. **xi + rho/tau** -- adds the blocked tau covariance (TAU_0_PLUS / TAU_2_PLUS, + with the tau_0<->tau_2 cross-correlation the truncated CosmoCov tau + covariance carries) and the verbatim RHO_STATS table. + +The rho/tau product needs the rho/tau *sidecar* HDUs: the RHO_STATS table carries +per-mode ``varrho_*`` variances the analysis SACC does not store, so the +converter copies them from the sidecar exactly as today's assembly does. A teeth +test pins that a perturbed input moves the output. + +The reference builder imports ``cosmosis_fitting.py`` by path (it is a script, +not a package module), skipping cleanly if a dependency is missing -- the same +loader pattern as ``test_cosmosis_fitting.py``. +""" + +import importlib.util +from pathlib import Path + +import numpy as np +import pytest +from astropy.io import fits + +from sp_validation import sacc_io, twopoint_convert + +_SCRIPT = ( + Path(__file__).resolve().parents[3] + / "cosmo_inference" + / "scripts" + / "cosmosis_fitting.py" +) + + +def _load_cf(): + """Import cosmosis_fitting.py by path; skip cleanly if a dep is missing.""" + if not _SCRIPT.exists(): + pytest.skip(f"cosmosis_fitting.py not found at {_SCRIPT}") + spec = importlib.util.spec_from_file_location("cosmosis_fitting", _SCRIPT) + module = importlib.util.module_from_spec(spec) + try: + spec.loader.exec_module(module) + except ImportError as exc: # pragma: no cover - container has numpy/astropy + pytest.importorskip(getattr(exc, "name", "") or "cosmosis_fitting_dependency") + raise + return module + + +cf = _load_cf() + + +# --- deterministic synthetic inputs ----------------------------------------- +# +# One tomographic bin (today's analysis). N_ANG angular bins on an ascending +# theta grid; N_ELL bandpowers; a 200-point n(z) on a uniform z grid (the DES +# NZDATA table needs a uniform Z_MID axis). The xi covariance is a full +# (2*N_ANG) matrix (xi+/xi- cross-block nonzero, as CosmoCov produces); the tau +# covariance is a full (3*N_ANG) matrix that the assembly truncates to its first +# 2 statistics (tau_0, tau_2). + +N_ANG = 5 +N_ELL = 8 +SOURCE = sacc_io.source_name(0) +PSF = sacc_io.PSF_TRACER + + +def _spd(n, seed): + """A symmetric positive-definite (n, n) matrix, seeded and recognizable.""" + rng = np.random.default_rng(seed) + a = rng.normal(size=(n, n)) + return a @ a.T + n * np.eye(n) + + +def _inputs(seed=0): + """Deterministic synthetic statistics + covariances for one bin pair.""" + rng = np.random.default_rng(seed) + theta = np.sort(rng.uniform(1.0, 250.0, N_ANG)) + ell = np.sort(rng.uniform(30.0, 3000.0, N_ELL)) + z = np.linspace(0.0125, 0.4875, 200) + return { + "theta": theta, + "ell": ell, + "z": z, + "nz": np.exp(-((z - 0.25) ** 2) / 0.02), + "xip": rng.uniform(1e-6, 1e-4, N_ANG), + "xim": rng.uniform(1e-6, 1e-4, N_ANG), + "cl_ee": rng.uniform(1e-10, 1e-8, N_ELL), + "cl_bb": rng.uniform(1e-11, 1e-9, N_ELL), + "cl_eb": rng.uniform(-1e-11, 1e-11, N_ELL), + "tau0p": rng.uniform(1e-6, 1e-5, N_ANG), + "tau2p": rng.uniform(1e-6, 1e-5, N_ANG), + "tau0m": rng.uniform(1e-6, 1e-5, N_ANG), + "tau2m": rng.uniform(1e-6, 1e-5, N_ANG), + "xi_cov": _spd(2 * N_ANG, seed + 1), + "cl_cov": _spd(N_ELL, seed + 2), + "tau_cov_full": _spd(3 * N_ANG, seed + 3), + } + + +# --- reference sidecar files + HDUs (the current script's own builders) ------ + + +def _rho_sidecar_hdu(theta, seed=7): + """A rho-stats BinTableHDU with the 25-column layout (values + variances).""" + rng = np.random.default_rng(seed) + columns = [fits.Column(name="theta", format="D", array=theta)] + for k in range(6): + for suffix in ("_p", "_m"): + columns.append( + fits.Column( + name=f"rho_{k}{suffix}", + format="D", + array=rng.uniform(-1e-3, 1e-3, len(theta)), + ) + ) + columns.append( + fits.Column( + name=f"varrho_{k}{suffix}", + format="D", + array=rng.uniform(1e-15, 1e-12, len(theta)), + ) + ) + return fits.BinTableHDU.from_columns(fits.ColDefs(columns)) + + +def _tau_sidecar_hdu(theta, tau0p, tau2p): + """A tau-stats BinTableHDU with theta + tau_0_p + tau_2_p columns.""" + columns = [ + fits.Column(name="theta", format="D", array=theta), + fits.Column(name="tau_0_p", format="D", array=tau0p), + fits.Column(name="tau_2_p", format="D", array=tau2p), + ] + return fits.BinTableHDU.from_columns(fits.ColDefs(columns)) + + +def _reference_fits(tmp_path, inp, *, cl=False, rho_tau=False): + """Build the reference 2pt-FITS with cosmosis_fitting.py's own functions. + + Reproduces the exact ``__main__`` HDU list for the requested configuration, + which is the assembly the converter must match byte for byte. + """ + nz_txt = tmp_path / "nz.txt" + np.savetxt(nz_txt, np.column_stack([inp["z"], inp["nz"]])) + cov_txt = tmp_path / "cov_xi.txt" + np.savetxt(cov_txt, inp["xi_cov"]) + + nz_hdu = cf.nz_to_fits(str(nz_txt)) + xip_hdu = cf._create_2pt_hdu(inp["xip"], inp["theta"], "XI_PLUS", "G+R", "G+R") + xim_hdu = cf._create_2pt_hdu(inp["xim"], inp["theta"], "XI_MINUS", "G-R", "G-R") + + hdu_list = [fits.PrimaryHDU(), nz_hdu] + + if rho_tau: + tau_cov_npy = tmp_path / "cov_tau.npy" + np.save(tau_cov_npy, inp["tau_cov_full"]) + cov_hdu = cf.covdat_to_fits(str(cov_txt), filename_cov_tau=str(tau_cov_npy)) + else: + cov_hdu = cf.covdat_to_fits(str(cov_txt), filename_cov_tau=None) + hdu_list.append(cov_hdu) + + if cl: + cl_block = np.zeros((5, N_ELL)) + cl_block[0], cl_block[1], cl_block[4] = inp["ell"], inp["cl_ee"], inp["cl_bb"] + cl_npy = tmp_path / "cl.npy" + np.save(cl_npy, cl_block) + cl_cov_npy = tmp_path / "cl_cov.npy" + np.save(cl_cov_npy, inp["cl_cov"]) + ell_r, cl_ee_r, cl_bb_r = cf.load_pseudo_cl(str(cl_npy)) + cl_ee_hdu, _cl_bb_hdu = cf.cl_to_fits(ell_r, cl_ee_r, cl_bb_r) + cov_cl_hdu = cf.cov_cl_to_fits(str(cl_cov_npy), cov_hdu="COVAR_FULL") + hdu_list.append(cov_cl_hdu) + + hdu_list.extend([xip_hdu, xim_hdu]) + if cl: + hdu_list.append(cl_ee_hdu) + + if rho_tau: + rho_path = tmp_path / "rho.fits" + fits.HDUList([fits.PrimaryHDU(), _rho_sidecar_hdu(inp["theta"])]).writeto( + rho_path, overwrite=True + ) + tau_path = tmp_path / "tau.fits" + fits.HDUList( + [ + fits.PrimaryHDU(), + _tau_sidecar_hdu(inp["theta"], inp["tau0p"], inp["tau2p"]), + ] + ).writeto(tau_path, overwrite=True) + rho_hdu = cf.rho_to_fits(str(rho_path), theta=inp["theta"]) + tau0_hdu, tau2_hdu = cf.tau_to_fits(str(tau_path), theta=inp["theta"]) + hdu_list.extend([tau0_hdu, tau2_hdu, rho_hdu]) + + out = tmp_path / "reference.fits" + fits.HDUList(hdu_list).writeto(out, overwrite=True) + return out + + +# --- the SACC each configuration is built from ------------------------------ + + +def _sacc(inp, *, cl=False, rho_tau=False): + """Build the analysis SACC the converter reads, matching ``_inputs``. + + The covariance is laid out to match the reference exactly: the xi block is + the full (2*N_ANG) matrix; the Cl block is the EE bandpower covariance; the + tau blocks carry the tau_0<->tau_2 cross-correlation from the truncated + CosmoCov tau covariance. Blocks not consumed by the 2pt-FITS (Cl BB/EB, tau + minus) get an identity block so ``add_covariance`` sees a full matrix. + """ + s = sacc_io.new_sacc({0: (inp["z"], inp["nz"])}) + sacc_io.add_xi(s, (0, 0), inp["theta"], inp["xip"], inp["xim"], grid="coarse") + if cl: + sacc_io.add_pseudo_cl( + s, + (0, 0), + inp["ell"], + inp["cl_ee"], + inp["cl_bb"], + inp["cl_eb"], + window_ells=np.arange(2, 102), + window_weights=np.random.default_rng(9).uniform(0, 1, (100, N_ELL)), + ) + if rho_tau: + sacc_io.add_tau(s, (0, 0), 0, inp["theta"], inp["tau0p"], inp["tau0m"]) + sacc_io.add_tau(s, (0, 0), 2, inp["theta"], inp["tau2p"], inp["tau2m"]) + + n = len(s.mean) + full = np.zeros((n, n)) + ip = s.indices(sacc_io.XI_PLUS, (SOURCE, SOURCE)) + im = s.indices(sacc_io.XI_MINUS, (SOURCE, SOURCE)) + xi_all = np.concatenate([ip, im]) + full[np.ix_(xi_all, xi_all)] = inp["xi_cov"] + + if cl: + iee = s.indices(sacc_io.CL_EE, (SOURCE, SOURCE)) + full[np.ix_(iee, iee)] = inp["cl_cov"] + for dtype in (sacc_io.CL_BB, sacc_io.CL_EB): + idx = s.indices(dtype, (SOURCE, SOURCE)) + full[np.ix_(idx, idx)] = np.eye(N_ELL) + if rho_tau: + t0p = s.indices(sacc_io.TAU_PLUS.format(k=0), (SOURCE, PSF)) + t2p = s.indices(sacc_io.TAU_PLUS.format(k=2), (SOURCE, PSF)) + tau_pp = np.concatenate([t0p, t2p]) + # The joint [tau_0+; tau_2+] block is the truncated CosmoCov tau + # covariance -- cross-correlation kept, matching covdat_to_fits. + full[np.ix_(tau_pp, tau_pp)] = inp["tau_cov_full"][: 2 * N_ANG, : 2 * N_ANG] + for dtype in (sacc_io.TAU_MINUS.format(k=0), sacc_io.TAU_MINUS.format(k=2)): + idx = s.indices(dtype, (SOURCE, PSF)) + full[np.ix_(idx, idx)] = np.eye(N_ANG) + s.add_covariance(full) + return s + + +def _sidecar_hdus(tmp_path, inp): + """Return the (rho_hdu, tau_hdu) sidecar input HDUs for the rho/tau product.""" + rho_path = tmp_path / "rho_in.fits" + fits.HDUList([fits.PrimaryHDU(), _rho_sidecar_hdu(inp["theta"])]).writeto( + rho_path, overwrite=True + ) + tau_path = tmp_path / "tau_in.fits" + fits.HDUList( + [fits.PrimaryHDU(), _tau_sidecar_hdu(inp["theta"], inp["tau0p"], inp["tau2p"])] + ).writeto(tau_path, overwrite=True) + with fits.open(rho_path) as r, fits.open(tau_path) as t: + return r[1].copy(), t[1].copy() + + +# ============================================================================= +# Byte-compare: the three product shapes +# ============================================================================= + + +def test_plain_xi_byte_equal(tmp_path): + """Plain-xi product: converter matches cosmosis_fitting.py byte for byte.""" + inp = _inputs(seed=0) + reference = _reference_fits(tmp_path, inp) + s = _sacc(inp) + out = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits(s, str(out), n_bins=1) + assert out.read_bytes() == reference.read_bytes() + + +def test_xi_cl_byte_equal(tmp_path): + """xi + pseudo-Cl product: COVMAT_CELL + CELL_EE reproduced byte for byte.""" + inp = _inputs(seed=10) + reference = _reference_fits(tmp_path, inp, cl=True) + s = _sacc(inp, cl=True) + out = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits(s, str(out), n_bins=1) + assert out.read_bytes() == reference.read_bytes() + + +def test_xi_rho_tau_byte_equal(tmp_path): + """xi + rho/tau product: blocked tau covariance + verbatim RHO_STATS match.""" + inp = _inputs(seed=20) + reference = _reference_fits(tmp_path, inp, rho_tau=True) + s = _sacc(inp, rho_tau=True) + rho_hdu, tau_hdu = _sidecar_hdus(tmp_path, inp) + out = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits( + s, str(out), rho_stats_hdu=rho_hdu, tau_stats_hdu=tau_hdu, n_bins=1 + ) + assert out.read_bytes() == reference.read_bytes() + + +# ============================================================================= +# Structural + teeth checks +# ============================================================================= + + +def test_tau_covariance_keeps_tau0_tau2_cross(tmp_path): + """The tau covariance block couples tau_0 and tau_2 (not block-diagonal). + + covdat_to_fits truncates the 3-statistic CosmoCov tau covariance to its + first 2 blocks and lays it in as ONE contiguous block, so tau_0<->tau_2 + cross-terms survive. A block-diagonal shortcut would zero them; pin that + the converter keeps them. + """ + inp = _inputs(seed=20) + s = _sacc(inp, rho_tau=True) + rho_hdu, tau_hdu = _sidecar_hdus(tmp_path, inp) + out = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits( + s, str(out), rho_stats_hdu=rho_hdu, tau_stats_hdu=tau_hdu, n_bins=1 + ) + with fits.open(out) as hdul: + cov = hdul["COVMAT"].data + # tau_0 block rows 2*N_ANG..3*N_ANG, tau_2 block 3*N_ANG..4*N_ANG. + cross = cov[2 * N_ANG : 3 * N_ANG, 3 * N_ANG : 4 * N_ANG] + expected = inp["tau_cov_full"][:N_ANG, N_ANG : 2 * N_ANG] + assert np.array_equal(cross, expected) + assert np.any(cross != 0.0) + + +def test_perturbed_xi_changes_output(tmp_path): + """Teeth: a changed xi+ input must change the converted data vector.""" + inp = _inputs(seed=0) + s = _sacc(inp) + out = tmp_path / "base.fits" + twopoint_convert.sacc_to_twopoint_fits(s, str(out), n_bins=1) + with fits.open(out) as hdul: + base_xip = hdul["XI_PLUS"].data["VALUE"].copy() + + inp2 = _inputs(seed=0) + inp2["xip"] = inp2["xip"] + 1.0 + s2 = _sacc(inp2) + out2 = tmp_path / "perturbed.fits" + twopoint_convert.sacc_to_twopoint_fits(s2, str(out2), n_bins=1) + with fits.open(out2) as hdul: + new_xip = hdul["XI_PLUS"].data["VALUE"] + + assert not np.array_equal(base_xip, new_xip) + assert np.array_equal(new_xip, inp2["xip"]) + + +def test_rho_tau_sidecars_required_together(tmp_path): + """Supplying only one of the rho/tau sidecars is a loud error.""" + inp = _inputs(seed=20) + s = _sacc(inp, rho_tau=True) + rho_hdu, _tau_hdu = _sidecar_hdus(tmp_path, inp) + with pytest.raises(ValueError, match="together"): + twopoint_convert.sacc_to_twopoint_fits( + s, str(tmp_path / "x.fits"), rho_stats_hdu=rho_hdu, n_bins=1 + ) diff --git a/src/sp_validation/tests/test_twopoint_convert_realdata.py b/src/sp_validation/tests/test_twopoint_convert_realdata.py new file mode 100644 index 00000000..9d2739a6 --- /dev/null +++ b/src/sp_validation/tests/test_twopoint_convert_realdata.py @@ -0,0 +1,268 @@ +"""Candide-local byte-compare of the converter against real 2pt-FITS products. + +Skipped unless a real product exists on disk (candide only; never committed). +It closes the loop end to end on real data: take a real CosmoSIS 2pt-FITS, build +an analysis SACC from its own contents, convert that SACC back to a 2pt-FITS, and +byte-compare. + +The reference is *not* the on-disk file directly. The committed on-disk products +were written by an older ``cosmosis_fitting.py`` (they carry a CELL_BB HDU and +order COVMAT before NZ_SOURCE); the converter reproduces the *current* script, +which the PR-4 migration will use to regenerate them. So the meaningful contract +-- "the converter equals the current writer" -- is tested by running the current +``cosmosis_fitting.py`` builders on the same real contents and byte-comparing the +converter against that. For transparency the test also records the direct diff +against the stale on-disk file, and asserts only that it differs *by whole HDUs* +(the extra CELL_BB), not in any shared block -- i.e. the drift is purely the +known HDU-set change, with no silent data corruption. + +Observed on 2026-07-10 for ``SP_v1.4.6_leak_corr`` and a ``glass_mock`` sibling: +converter == current-script byte for byte; converter vs on-disk differs only by +the CELL_BB HDU. +""" + +import importlib.util +from pathlib import Path + +import numpy as np +import pytest +from astropy.io import fits + +from sp_validation import sacc_io, twopoint_convert + +_DATA = Path("/automnt/n17data/cdaley/unions/code/sp_validation/cosmo_inference/data") +_REAL_FILES = { + "SP_v1.4.6_leak_corr": _DATA + / "SP_v1.4.6_leak_corr_A_minsep=1.0_maxsep=250.0_nbins=20_npatch=1" + / "cosmosis_SP_v1.4.6_leak_corr_A_minsep=1.0_maxsep=250.0_nbins=20_npatch=1.fits", + "glass_mock_00001": _DATA / "glass_mock_00001" / "cosmosis_glass_mock_00001.fits", +} + +_SCRIPT = ( + Path(__file__).resolve().parents[3] + / "cosmo_inference" + / "scripts" + / "cosmosis_fitting.py" +) + + +def _load_cf(): + spec = importlib.util.spec_from_file_location("cosmosis_fitting", _SCRIPT) + module = importlib.util.module_from_spec(spec) + spec.loader.exec_module(module) + return module + + +def _sacc_from_2pt_fits(hdul): + """Build an analysis SACC from a real CosmoSIS 2pt-FITS's own contents. + + Reads the NZ, ξ±, pseudo-Cℓ (EE/BB) and blocked covariance back out of the + product and lays them into the standard SACC layout — the inverse direction + the converter then undoes. τ± minus and Cℓ EB are not in the 2pt-FITS, so + they are stored as zeros with identity covariance sub-blocks (the converter + consumes only the ``+``/EE parts). + """ + z = hdul["NZ_SOURCE"].data["Z_MID"].astype(float) + nz = hdul["NZ_SOURCE"].data["BIN1"].astype(float) + theta = hdul["XI_PLUS"].data["ANG"].astype(float) + xip = hdul["XI_PLUS"].data["VALUE"].astype(float) + xim = hdul["XI_MINUS"].data["VALUE"].astype(float) + n = len(theta) + + ell = hdul["CELL_EE"].data["ANG"].astype(float) + cl_ee = hdul["CELL_EE"].data["VALUE"].astype(float) + cl_bb = hdul["CELL_BB"].data["VALUE"].astype(float) + n_ell = len(ell) + + covmat = hdul["COVMAT"].data.astype(float) + covmat_cell = hdul["COVMAT_CELL"].data.astype(float) + xi_cov = covmat[: 2 * n, : 2 * n] + tau_joint = covmat[2 * n : 4 * n, 2 * n : 4 * n] + + tau0p = hdul["TAU_0_PLUS"].data["VALUE"].astype(float) + tau2p = hdul["TAU_2_PLUS"].data["VALUE"].astype(float) + + s = sacc_io.new_sacc({0: (z, nz)}) + sacc_io.add_xi(s, (0, 0), theta, xip, xim, grid="coarse") + sacc_io.add_pseudo_cl( + s, + (0, 0), + ell, + cl_ee, + cl_bb, + np.zeros(n_ell), + window_ells=np.arange(2, 102), + window_weights=np.ones((100, n_ell)), + ) + sacc_io.add_tau(s, (0, 0), 0, theta, tau0p, np.zeros(n)) + sacc_io.add_tau(s, (0, 0), 2, theta, tau2p, np.zeros(n)) + + source, psf = sacc_io.source_name(0), sacc_io.PSF_TRACER + idx = { + "xi": np.concatenate( + [ + s.indices(sacc_io.XI_PLUS, (source, source)), + s.indices(sacc_io.XI_MINUS, (source, source)), + ] + ), + "ee": s.indices(sacc_io.CL_EE, (source, source)), + "bb": s.indices(sacc_io.CL_BB, (source, source)), + "eb": s.indices(sacc_io.CL_EB, (source, source)), + "t0p": s.indices(sacc_io.TAU_PLUS.format(k=0), (source, psf)), + "t0m": s.indices(sacc_io.TAU_MINUS.format(k=0), (source, psf)), + "t2p": s.indices(sacc_io.TAU_PLUS.format(k=2), (source, psf)), + "t2m": s.indices(sacc_io.TAU_MINUS.format(k=2), (source, psf)), + } + full = np.zeros((len(s.mean), len(s.mean))) + full[np.ix_(idx["xi"], idx["xi"])] = xi_cov + full[np.ix_(idx["ee"], idx["ee"])] = covmat_cell + tau_pp = np.concatenate([idx["t0p"], idx["t2p"]]) + full[np.ix_(tau_pp, tau_pp)] = tau_joint + for key in ("bb", "eb"): + full[np.ix_(idx[key], idx[key])] = np.eye(n_ell) + for key in ("t0m", "t2m"): + full[np.ix_(idx[key], idx[key])] = np.eye(n) + s.add_covariance(full) + + tau_sidecar = fits.BinTableHDU.from_columns( + fits.ColDefs( + [ + fits.Column(name="theta", format="D", array=theta), + fits.Column(name="tau_0_p", format="D", array=tau0p), + fits.Column(name="tau_2_p", format="D", array=tau2p), + ] + ) + ) + return s, hdul["RHO_STATS"].copy(), tau_sidecar + + +def _current_script_reference(cf, hdul, tmp_path): + """Reference: run the current cosmosis_fitting.py on the real file's contents.""" + z = hdul["NZ_SOURCE"].data["Z_MID"].astype(float) + nz = hdul["NZ_SOURCE"].data["BIN1"].astype(float) + theta = hdul["XI_PLUS"].data["ANG"].astype(float) + xip = hdul["XI_PLUS"].data["VALUE"].astype(float) + xim = hdul["XI_MINUS"].data["VALUE"].astype(float) + n = len(theta) + ell = hdul["CELL_EE"].data["ANG"].astype(float) + cl_ee = hdul["CELL_EE"].data["VALUE"].astype(float) + cl_bb = hdul["CELL_BB"].data["VALUE"].astype(float) + covmat = hdul["COVMAT"].data.astype(float) + covmat_cell = hdul["COVMAT_CELL"].data.astype(float) + tau0p = hdul["TAU_0_PLUS"].data["VALUE"].astype(float) + tau2p = hdul["TAU_2_PLUS"].data["VALUE"].astype(float) + + np.savetxt(tmp_path / "nz.txt", np.column_stack([z, nz])) + np.savetxt(tmp_path / "cov.txt", covmat[: 2 * n, : 2 * n]) + tau_cov = np.zeros((3 * n, 3 * n)) + tau_cov[: 2 * n, : 2 * n] = covmat[2 * n : 4 * n, 2 * n : 4 * n] + np.save(tmp_path / "cov_tau.npy", tau_cov) + cl_block = np.zeros((5, len(ell))) + cl_block[0], cl_block[1], cl_block[4] = ell, cl_ee, cl_bb + np.save(tmp_path / "cl.npy", cl_block) + np.save(tmp_path / "cl_cov.npy", covmat_cell) + + nz_hdu = cf.nz_to_fits(str(tmp_path / "nz.txt")) + xip_hdu = cf._create_2pt_hdu(xip, theta, "XI_PLUS", "G+R", "G+R") + xim_hdu = cf._create_2pt_hdu(xim, theta, "XI_MINUS", "G-R", "G-R") + cov_hdu = cf.covdat_to_fits( + str(tmp_path / "cov.txt"), filename_cov_tau=str(tmp_path / "cov_tau.npy") + ) + ell_r, cl_ee_r, cl_bb_r = cf.load_pseudo_cl(str(tmp_path / "cl.npy")) + cl_ee_hdu, _ = cf.cl_to_fits(ell_r, cl_ee_r, cl_bb_r) + cov_cl_hdu = cf.cov_cl_to_fits(str(tmp_path / "cl_cov.npy"), cov_hdu="COVAR_FULL") + + fits.HDUList([fits.PrimaryHDU(), hdul["RHO_STATS"].copy()]).writeto( + tmp_path / "rho.fits", overwrite=True + ) + rho_hdu = cf.rho_to_fits(str(tmp_path / "rho.fits"), theta=theta) + tau_sidecar = fits.BinTableHDU.from_columns( + fits.ColDefs( + [ + fits.Column(name="theta", format="D", array=theta), + fits.Column(name="tau_0_p", format="D", array=tau0p), + fits.Column(name="tau_2_p", format="D", array=tau2p), + ] + ) + ) + fits.HDUList([fits.PrimaryHDU(), tau_sidecar]).writeto( + tmp_path / "tau.fits", overwrite=True + ) + tau0_hdu, tau2_hdu = cf.tau_to_fits(str(tmp_path / "tau.fits"), theta=theta) + + out = tmp_path / "reference.fits" + fits.HDUList( + [ + fits.PrimaryHDU(), + nz_hdu, + cov_hdu, + cov_cl_hdu, + xip_hdu, + xim_hdu, + cl_ee_hdu, + tau0_hdu, + tau2_hdu, + rho_hdu, + ] + ).writeto(out, overwrite=True) + return out + + +@pytest.mark.parametrize("label", list(_REAL_FILES)) +def test_realdata_roundtrip_byte_equal(label, tmp_path): + """Converter reproduces the current writer byte for byte on a real product.""" + real = _REAL_FILES[label] + if not real.exists(): + pytest.skip(f"real 2pt-FITS not on disk: {real}") + cf = _load_cf() + + with fits.open(real) as hdul: + s, rho_hdu, tau_hdu = _sacc_from_2pt_fits(hdul) + reference = _current_script_reference(cf, hdul, tmp_path) + + converted = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits( + s, str(converted), rho_stats_hdu=rho_hdu, tau_stats_hdu=tau_hdu, n_bins=1 + ) + + # The contract: converter == current cosmosis_fitting.py, byte for byte. + assert converted.read_bytes() == reference.read_bytes() + + +@pytest.mark.parametrize("label", list(_REAL_FILES)) +def test_realdata_ondisk_drift_is_only_cell_bb(label, tmp_path): + """The stale on-disk file differs from the converter *only* by the CELL_BB HDU. + + Documents (and guards) the known script-version drift: the on-disk products + were written before CELL_BB was dropped from the assembly, so they carry one + extra HDU. Every HDU the two share carries the same data to floating-point + precision — the drift is a whole-HDU addition, never a silent change to a + shared block. (Bin-edge columns differ by ~1e-17 float noise: the stale file + stored a clean ``Z_LOW=0``, while ``z_mid - step/2`` rounds to ``-1.7e-18``; + both are the same number, so the shared-data check is ``allclose``, not + bitwise — bitwise equality is asserted against the *current* writer above.) + """ + real = _REAL_FILES[label] + if not real.exists(): + pytest.skip(f"real 2pt-FITS not on disk: {real}") + + with fits.open(real) as hdul: + s, rho_hdu, tau_hdu = _sacc_from_2pt_fits(hdul) + ondisk_names = [h.name for h in hdul] + converted = tmp_path / "converted.fits" + twopoint_convert.sacc_to_twopoint_fits( + s, str(converted), rho_stats_hdu=rho_hdu, tau_stats_hdu=tau_hdu, n_bins=1 + ) + with fits.open(converted) as conv: + conv_names = [h.name for h in conv] + # The only HDU the on-disk file has that the converter does not. + assert set(ondisk_names) - set(conv_names) == {"CELL_BB"} + # Every shared table HDU carries the same data to float precision. + for name in conv_names: + if name == "PRIMARY": + continue + a, b = hdul[name].data, conv[name].data + if hasattr(a, "names"): + assert all(np.allclose(a[c], b[c]) for c in a.names), name + else: + assert np.allclose(a, b), name diff --git a/src/sp_validation/twopoint_convert.py b/src/sp_validation/twopoint_convert.py new file mode 100644 index 00000000..c06e0eb3 --- /dev/null +++ b/src/sp_validation/twopoint_convert.py @@ -0,0 +1,330 @@ +"""TWOPOINT_CONVERT. + +:Name: twopoint_convert.py + +:Description: Convert an analysis SACC file into the "2pt FITS" that CosmoSIS's + ``2pt_like`` (and Sacha Guerrini's ρ/τ ``2pt_like_xi_sys`` fork) + reads. The output reproduces today's hand-assembled product from + ``cosmo_inference/scripts/cosmosis_fitting.py`` HDU-for-HDU and + byte-for-byte: an NZDATA table, XI_PLUS / XI_MINUS 2pt tables, + optional CELL_EE / CELL_BB pseudo-Cℓ tables, the blocked COVMAT + (with ``STRT_i`` block-offset headers) and separate COVMAT_CELL, + and — when the SACC carries them — the TAU_{0,2}_PLUS 2pt tables + and the RHO_STATS table. + + The converter is the *inverse* of the SACC writers in + :mod:`sp_validation.sacc_io`: it reads statistics back through + those readers and lays them into the DES ``twopoint`` FITS + convention. SACC's canonical order is pair-major (per pair + ``[ξ+; ξ−]``); the 2pt-FITS layout is type-major (all ξ+, then all + ξ−), so the data-vector and its covariance are permuted here via + ``s.indices`` rather than assuming any global order. + + Scope note (single-bin today, tomography-ready): the assembly this + mirrors is single-tomographic-bin — BIN1/BIN2 are all 1, one NZ + ``BIN1`` column. The converter reads bin ``(0, 0)`` accordingly. + A tomographic 2pt-FITS layout (multiple bin pairs, per-pair + BIN1/BIN2, one NZ column per bin) is a later extension; it is not + what today's CosmoSIS pipeline consumes, so it is out of scope for + the byte-compatible converter. + + Rho/tau caveat: the SACC layout stores ρ±/τ± *values* only, while + the 2pt-FITS RHO_STATS table also carries the per-mode *variances* + (``varrho_*``) that Sacha's fork's covariance path reads. Those + variances are not recoverable from the analysis SACC. The + converter therefore writes RHO_STATS / TAU HDUs only when a + ``rho_stats``/``tau_stats`` sidecar FITS is supplied (the same + file today's assembly copies verbatim); it never fabricates + variances. ξ±, Cℓ, n(z) and the covariance — the data vector + CosmoSIS fits — are fully reconstructed from SACC alone. +""" + +import numpy as np +from astropy.io import fits + +from . import sacc_io + +# The QUANT1/QUANT2 header pair CosmoSIS stamps on each 2pt table, keyed by the +# extension name — copied from cosmosis_fitting.py so the headers match card +# for card. +_QUANT = { + "XI_PLUS": ("G+R", "G+R"), + "XI_MINUS": ("G-R", "G-R"), + "CELL_EE": ("GEF", "GEF"), + "CELL_BB": ("GBF", "GBF"), + "TAU_0_PLUS": ("G+R", "P+R"), + "TAU_2_PLUS": ("G+R", "SR+R"), +} + + +def _twopoint_hdu(name, values, ang, *, ang_unit=None): + """Build one 2pt BinTableHDU (BIN1/BIN2/ANGBIN/VALUE/ANG). + + Reproduces ``cosmosis_fitting.py._create_2pt_hdu`` /``cl_to_fits`` exactly: + same column order and formats, the ``2PTDATA`` marker, the QUANT pair for + ``name``, and NZ_SOURCE kernels. ``ang_unit`` stamps ``TUNIT`` on the ANG + column ("arcmin" for real-space ξ/τ; unset for Cℓ, whose ANG is ℓ). + """ + nbins = len(values) + angbin = np.arange(1, nbins + 1) + columns = [ + fits.Column(name="BIN1", format="K", array=np.ones(nbins)), + fits.Column(name="BIN2", format="K", array=np.ones(nbins)), + fits.Column(name="ANGBIN", format="K", array=angbin), + fits.Column(name="VALUE", format="D", array=values), + fits.Column(name="ANG", format="D", unit=ang_unit, array=ang), + ] + hdu = fits.BinTableHDU.from_columns(fits.ColDefs(columns), name=name) + quant1, quant2 = _QUANT[name] + for key, value in { + "2PTDATA": "T", + "QUANT1": quant1, + "QUANT2": quant2, + "KERNEL_1": "NZ_SOURCE", + "KERNEL_2": "NZ_SOURCE", + "WINDOWS": "SAMPLE", + }.items(): + hdu.header[key] = value + return hdu + + +def _nz_hdu(s, n_bins): + """Build the NZDATA HDU from the SACC ``source_i`` NZ tracers. + + Reproduces ``cosmosis_fitting.py.nz_to_fits``: Z_MID from the tracer ``z`` + grid (assumed uniform), Z_LOW/Z_HIGH as ± half a step, one ``BIN{i+1}`` + column per source bin, and the NZDATA/NBIN/NZ header cards. All source bins + are required to share the ``z`` grid — the single ``Z_MID`` axis of the + DES NZDATA table. + """ + z_mid, nz0 = sacc_io.get_nz(s, 0) + z_mid = np.asarray(z_mid, dtype=float) + step = z_mid[1] - z_mid[0] + z_low = z_mid - step / 2 + z_high = z_mid + step / 2 + + columns = [ + fits.Column(name="Z_LOW", format="D", array=z_low), + fits.Column(name="Z_MID", format="D", array=z_mid), + fits.Column(name="Z_HIGH", format="D", array=z_high), + ] + for i in range(n_bins): + z_i, nz_i = sacc_io.get_nz(s, i) + if not np.array_equal(np.asarray(z_i, dtype=float), z_mid): + raise ValueError( + f"source bin {i} n(z) grid differs from source bin 0; the DES " + "NZDATA table requires one shared Z_MID axis" + ) + columns.append(fits.Column(name=f"BIN{i + 1}", format="D", array=nz_i)) + + hdu = fits.BinTableHDU.from_columns(fits.ColDefs(columns), name="NZDATA") + for key, value in { + "NZDATA": "T ", + "EXTNAME": "NZ_SOURCE", + "NBIN": n_bins, + "NZ": len(z_low), + }.items(): + hdu.header[key] = value + return hdu + + +def _cov_hdu(matrix, block_names, block_starts, extname="COVMAT", name_in_ctor=False): + """Build a covariance ImageHDU with ``NAME_i``/``STRT_i`` block headers. + + Reproduces the two covariance builders in ``cosmosis_fitting.py`` card for + card. The blocked ξ/τ ``covdat_to_fits`` builds ``ImageHDU(cov)`` unnamed + and stamps ``COVDATA`` then ``EXTNAME`` from a dict; the ``cov_cl_to_fits`` + CELL covariance builds ``ImageHDU(cov, name="COVMAT_CELL")`` (so the EXTNAME + card is created early, with astropy's standard comment) before re-stamping. + ``name_in_ctor`` selects the second form so the card order matches exactly. + """ + matrix = np.asarray(matrix, dtype=np.float64) + if matrix.shape[0] != matrix.shape[1]: + raise ValueError(f"covariance must be square; got shape {matrix.shape}") + hdu = fits.ImageHDU(matrix, name=extname) if name_in_ctor else fits.ImageHDU(matrix) + hdu.header["COVDATA"] = "True" + hdu.header["EXTNAME"] = extname + for i, (name, start) in enumerate(zip(block_names, block_starts)): + hdu.header[f"NAME_{i}"] = name + hdu.header[f"STRT_{i}"] = int(start) + return hdu + + +def _type_major_xi(s, bins): + """Return ``(theta, xip, xim)`` for one bin pair from the SACC coarse grid. + + ``sacc_io.get_xi`` already returns each statistic in insertion (= ascending + θ) order; the type-major split (all ξ+, then all ξ−) is exactly the two + arrays it hands back, so no further permutation is needed for a single pair. + """ + return sacc_io.get_xi(s, bins, grid="coarse") + + +def sacc_to_twopoint_fits( + s, + path, + *, + rho_stats_hdu=None, + tau_stats_hdu=None, + n_bins=1, +): + """Convert an analysis SACC to a CosmoSIS 2pt-FITS file. + + The assembled ``HDUList`` matches today's ``cosmosis_fitting.py`` product + for the configuration the SACC describes: PRIMARY, NZ_SOURCE, COVMAT, then + (if present) COVMAT_CELL, XI_PLUS, XI_MINUS, (if present) CELL_EE / CELL_BB, + and (if the rho/tau sidecars are supplied) TAU_0_PLUS, TAU_2_PLUS, + RHO_STATS. The data vector and its covariance are laid out type-major + (all ξ+, then all ξ−, then the τ blocks), which is the DES ``twopoint`` + convention CosmoSIS reads. + + Parameters + ---------- + s : sacc.Sacc + Analysis SACC (coarse ξ±, optional pseudo-Cℓ, covariance, and — for the + ρ/τ product — the τ data points; see ``rho_stats_hdu``). + path : str + Output FITS path (overwritten). + rho_stats_hdu, tau_stats_hdu : astropy.io.fits.BinTableHDU, optional + The rho-stats / tau-stats sidecar HDUs, copied verbatim as today's + assembly does. Required together to write the ρ/τ product; the SACC + alone cannot rebuild the ``varrho_*`` columns Sacha's fork reads. When + omitted, a pure ξ (± Cℓ) product is written. + n_bins : int, optional + Number of source tomographic bins (default 1, the current single-bin + analysis). Sets the NZDATA column count. + + Returns + ------- + astropy.io.fits.HDUList + The assembled list, also written to ``path``. + """ + if (rho_stats_hdu is None) != (tau_stats_hdu is None): + raise ValueError( + "rho_stats_hdu and tau_stats_hdu must be supplied together " + "(the ρ/τ product needs both, or neither for a pure-ξ product)" + ) + use_rho_tau = rho_stats_hdu is not None + bins = (0, 0) + + nz_hdu = _nz_hdu(s, n_bins) + theta, xip, xim = _type_major_xi(s, bins) + xip_hdu = _twopoint_hdu("XI_PLUS", xip, theta, ang_unit="arcmin") + xim_hdu = _twopoint_hdu("XI_MINUS", xim, theta, ang_unit="arcmin") + + cell_hdu, cov_cell_hdu = _build_cell(s, bins) + + cov_hdu = _build_covmat(s, bins, use_rho_tau=use_rho_tau) + + tau_hdus, rho_hdu = _build_rho_tau(rho_stats_hdu, tau_stats_hdu, theta, use_rho_tau) + + # HDU order mirrors cosmosis_fitting.py's __main__: PRIMARY, NZ, COVMAT, + # COVMAT_CELL, XI±, CELL_EE, then the τ/ρ tables. + hdu_list = [fits.PrimaryHDU(), nz_hdu, cov_hdu] + if cov_cell_hdu is not None: + hdu_list.append(cov_cell_hdu) + hdu_list.extend([xip_hdu, xim_hdu]) + if cell_hdu is not None: + hdu_list.append(cell_hdu) + if use_rho_tau: + hdu_list.extend([*tau_hdus, rho_hdu]) + + hdul = fits.HDUList(hdu_list) + hdul.writeto(path, overwrite=True) + return hdul + + +def _build_cell(s, bins): + """Build the CELL_EE 2pt HDU plus the COVMAT_CELL HDU from the SACC pseudo-Cℓ. + + Returns ``(None, None)`` when the SACC has no pseudo-Cℓ. Only CELL_EE is + emitted — the harmonic ``2pt_like`` fits ``data_sets=CELL_EE``, and today's + assembly appends CELL_EE alone (it builds a CELL_BB HDU but discards it). + The SACC still carries EE/BB/EB with bandpower windows for the B-mode + null-test path; this converter surfaces only the block CosmoSIS reads. The + CELL covariance (the EE bandpower covariance) lives in its own COVMAT_CELL + ImageHDU, matching today's product. + """ + if sacc_io.CL_EE not in s.get_data_types(): + return None, None + + ell, cl_ee, _cl_bb, _cl_eb, _window = sacc_io.get_pseudo_cl(s, bins) + cell_hdu = _twopoint_hdu("CELL_EE", cl_ee, ell) + cell_idx = s.indices(sacc_io.CL_EE, sacc_io._pair(bins)) + cov_cell = s.covariance.dense[np.ix_(cell_idx, cell_idx)] + cov_cell_hdu = _cov_hdu( + cov_cell, ["CELL_EE"], [0], extname="COVMAT_CELL", name_in_ctor=True + ) + return cell_hdu, cov_cell_hdu + + +def _build_covmat(s, bins, *, use_rho_tau): + """Assemble the blocked COVMAT (ξ± type-major, then the τ blocks). + + The ξ covariance is pulled from the SACC as the contiguous ξ+/ξ− block for + the pair and permuted from pair-major (SACC) to type-major (2pt-FITS). Under + ``use_rho_tau`` the τ_0/τ_2 covariance blocks are appended block-diagonally + with zero ξ↔τ cross-blocks, exactly as ``covdat_to_fits`` builds them. + """ + pair = sacc_io._pair(bins) + idx_p = s.indices(sacc_io.XI_PLUS, pair) + idx_m = s.indices(sacc_io.XI_MINUS, pair) + n_theta = len(idx_p) + xi_idx = np.concatenate([idx_p, idx_m]) # type-major permutation + xi_cov = s.covariance.dense[np.ix_(xi_idx, xi_idx)] + + names = ["XI_PLUS", "XI_MINUS"] + starts = [0, n_theta] + matrix = xi_cov + + if use_rho_tau: + # The τ covariance couples τ_0+ and τ_2+ (today's assembly truncates the + # 3-statistic CosmoCov τ covariance to its first 2 blocks and lays it in + # as ONE contiguous [τ_0+; τ_2+] block — cross-correlation kept). In the + # SACC those two selections are not adjacent (τ_0− sits between them), so + # gather both index sets and extract the joint sub-block, ξ↔τ zero. + tau_pair = (sacc_io.source_name(0), sacc_io.PSF_TRACER) + idx_tau0 = s.indices(sacc_io.TAU_PLUS.format(k=0), tau_pair) + idx_tau2 = s.indices(sacc_io.TAU_PLUS.format(k=2), tau_pair) + tau_idx = np.concatenate([idx_tau0, idx_tau2]) + tau_cov = s.covariance.dense[np.ix_(tau_idx, tau_idx)] + matrix = _block_diag(matrix, tau_cov) + names += ["TAU_0_PLUS", "TAU_2_PLUS"] + starts += [2 * n_theta, 2 * n_theta + len(idx_tau0)] + + return _cov_hdu(matrix, names, starts) + + +def _block_diag(*blocks): + """Stack square blocks block-diagonally with zero cross-blocks.""" + sizes = [b.shape[0] for b in blocks] + n = sum(sizes) + out = np.zeros((n, n)) + start = 0 + for b in blocks: + out[start : start + b.shape[0], start : start + b.shape[0]] = b + start += b.shape[0] + return out + + +def _build_rho_tau(rho_stats_hdu, tau_stats_hdu, theta, use_rho_tau): + """Build the TAU_{0,2}_PLUS 2pt HDUs and the verbatim RHO_STATS HDU. + + Mirrors ``tau_to_fits`` / ``rho_to_fits``: τ_0/τ_2 read their ``tau_k_p`` + columns onto the shared ξ θ grid (consistency step); RHO_STATS is copied + verbatim from the sidecar with its θ column forced onto the ξ grid. The + ``varrho_*`` columns ride along in the copy — they are why the sidecar is + required (the SACC cannot supply them). + """ + if not use_rho_tau: + return (), None + + tau = tau_stats_hdu.data + tau0_hdu = _twopoint_hdu("TAU_0_PLUS", tau["tau_0_p"], theta, ang_unit="arcmin") + tau2_hdu = _twopoint_hdu("TAU_2_PLUS", tau["tau_2_p"], theta, ang_unit="arcmin") + + rho_hdu = rho_stats_hdu.copy() + rho_hdu.name = "RHO_STATS" + rho_hdu.data = rho_hdu.data.copy() + rho_hdu.data["theta"] = theta + return (tau0_hdu, tau2_hdu), rho_hdu From f9b1d12199c414246ee8b67bf9b079bec389ff00 Mon Sep 17 00:00:00 2001 From: Cail Daley Date: Fri, 10 Jul 2026 03:27:10 +0200 Subject: [PATCH 2/3] feat(one_covariance): SACC <-> OneCovariance file-format glue Add sp_validation.one_covariance_io, the two small pieces of glue between the SACC layout and OneCovariance (which knows nothing of SACC and only exchanges files): - write_nz / nz_table / nz_config_stanza: the source_i NZ tracers of an analysis SACC -> OneCovariance's combined whitespace n(z) file (column 0 = shared z grid, one n(z) column per tomographic bin, no bin edges) plus the [redshift] config stanza pointing at it. Config keys verified against upstream OneCovariance config.ini (zlens_directory/zlens_file/value_loc_in_lensbin); the UNIONS template (cosmo_val/pseudo_cl.py._modify_onecov_config) uses the z_directory alias, exposed via dir_key. Fails fast on missing bins or disagreeing z grids. - covariance_blocks: OneCovariance's flat covariance_list_*.dat output -> dense square block(s) paired with SACC selectors for sacc_io.assemble_covariance, reusing statistics.cov_from_one_covariance for the per-block reshape (col 10 gaussian / col 9 gauss+non-gaussian). Single-block and multi-block (tomography-ready) forms. Tests (test_one_covariance_io.py, 9 passing): reshape to a hand-built matrix + gaussian/gauss+ng column selection + perturbation teeth; blocks feed assemble_covariance and round-trip s.covariance.dense; n(z) write/read round-trip, UNIONS dir_key, header/no-header, and fail-fast on mismatched grids / missing bin / bad value_loc. Co-Authored-By: Claude Fable 5 --- src/sp_validation/one_covariance_io.py | 256 ++++++++++++++++ .../tests/test_one_covariance_io.py | 274 ++++++++++++++++++ 2 files changed, 530 insertions(+) create mode 100644 src/sp_validation/one_covariance_io.py create mode 100644 src/sp_validation/tests/test_one_covariance_io.py diff --git a/src/sp_validation/one_covariance_io.py b/src/sp_validation/one_covariance_io.py new file mode 100644 index 00000000..4c0f2ae4 --- /dev/null +++ b/src/sp_validation/one_covariance_io.py @@ -0,0 +1,256 @@ +"""ONE_COVARIANCE_IO. + +:Name: one_covariance_io.py + +:Description: File-format glue between the SACC data-product layout + (:mod:`sp_validation.sacc_io`) and OneCovariance + (https://github.com/rreischke/OneCovariance). Two directions: + + - **n(z) SACC -> OneCovariance input** (:func:`write_nz`): the + ``source_i`` NZ tracers of an analysis SACC are written as the + combined whitespace-delimited redshift file OneCovariance reads + (column 0 = z grid, then one ``n(z)`` column per tomographic + bin, no bin edges), and a matching ``[redshift]`` config stanza + is returned via :func:`nz_config_stanza`. + + - **OneCovariance output -> SACC covariance blocks** + (:func:`covariance_blocks`): the flat ``covariance_list_*.dat`` + table OneCovariance emits (one row per element pair) is reshaped + into dense square block(s) — reusing + :func:`sp_validation.statistics.cov_from_one_covariance` for the + per-block reshape — and paired with SACC selectors so a caller + can feed them straight to + :func:`sp_validation.sacc_io.assemble_covariance`. + + OneCovariance itself is *not* a dependency: this module only + touches its file formats, verified against the upstream + ``config.ini`` (``rreischke/OneCovariance`` @ main). + + n(z) file format (upstream ``config.ini`` comment, verbatim): + + ``redshift n_1(z) ... n_{N_source}(z)`` + + i.e. a plain whitespace-delimited text file, column 0 the shared + redshift grid and one column per tomographic bin — no ``z_low``/ + ``z_high`` edges (this is the OneCovariance convention, distinct + from the CosmoSIS NZDATA table which *does* carry edges). All + source bins must therefore share one z grid. + + ``[redshift]`` config keys (upstream canonical names): a single + combined file goes in ``zlens_directory`` + ``zlens_file``; + ``value_loc_in_lensbin`` (``mid``/``left``/``right``) says where + in each histogram bin the tabulated ``n(z)`` value sits — ``mid`` + for the bin-centred grids the SACC stores. NOTE: the UNIONS + OneCovariance template driven by + ``cosmo_val/pseudo_cl.py._modify_onecov_config`` writes the older + key names ``z_directory``/``zlens_file`` instead; pass + ``dir_key="z_directory"`` to match that template. +""" + +import os + +import numpy as np + +from . import sacc_io +from .statistics import cov_from_one_covariance + + +def nz_table(s, n_bins): + """Stack the SACC ``source_i`` NZ tracers into a OneCovariance n(z) table. + + Parameters + ---------- + s : sacc.Sacc + SACC holding ``source_0 … source_{n_bins-1}`` NZ tracers. + n_bins : int + Number of tomographic source bins to write. + + Returns + ------- + numpy.ndarray + Array of shape ``(n_z, n_bins + 1)``: column 0 the shared redshift + grid, columns ``1 … n_bins`` the per-bin ``n(z)``. This is the + OneCovariance combined-file layout (``redshift n_1(z) … n_N(z)``). + + Raises + ------ + ValueError + If any source bin is missing, or if the bins do not share one z grid + (OneCovariance's combined file has a single redshift column, so the + grids must agree bin-for-bin). + """ + z0, nz0 = sacc_io.get_nz(s, 0) + z0 = np.asarray(z0, dtype=float) + columns = [z0] + for i in range(n_bins): + if sacc_io.source_name(i) not in s.tracers: + raise ValueError( + f"SACC has no NZ tracer {sacc_io.source_name(i)!r}; cannot write " + f"a {n_bins}-bin OneCovariance n(z) file" + ) + z_i, nz_i = sacc_io.get_nz(s, i) + if not np.array_equal(np.asarray(z_i, dtype=float), z0): + raise ValueError( + f"source bin {i} n(z) grid differs from source bin 0; the " + "OneCovariance combined n(z) file has one shared redshift column" + ) + columns.append(np.asarray(nz_i, dtype=float)) + return np.column_stack(columns) + + +def write_nz(s, path, n_bins, *, dir_key="zlens_directory", header=True): + """Write the OneCovariance combined n(z) input file from a SACC. + + OneCovariance reads the source redshift distribution as a plain + whitespace-delimited text file whose column 0 is the shared redshift grid + and whose remaining columns are the per-bin ``n(z)`` (``redshift n_1(z) + … n_N(z)``) — no ``z_low``/``z_high`` edges. This writes that file from the + SACC ``source_i`` NZ tracers and returns the ``[redshift]`` config stanza + that points OneCovariance at it. + + Parameters + ---------- + s : sacc.Sacc + Analysis SACC with the ``source_i`` NZ tracers. + path : str or pathlib.Path + Output text-file path (overwritten). Its directory + basename become + the ``[redshift]`` directory/file config values. + n_bins : int + Number of tomographic source bins to write. + dir_key : str, optional + Config key for the redshift directory. Default ``"zlens_directory"`` + (upstream canonical). Pass ``"z_directory"`` for the UNIONS template + driven by ``pseudo_cl.py._modify_onecov_config``. + header : bool, optional + If ``True`` (default) prepend a ``# redshift n_1(z) …`` comment header + naming the columns; OneCovariance's ``genfromtxt``-style reader ignores + it. Set ``False`` for a bare numeric file. + + Returns + ------- + dict + The ``[redshift]`` config stanza (see :func:`nz_config_stanza`), naming + the file just written. + """ + table = nz_table(s, n_bins) + head = "" + if header: + cols = " ".join(f"n_{i + 1}(z)" for i in range(n_bins)) + head = f"redshift {cols}" + np.savetxt(str(path), table, header=head) + return nz_config_stanza( + os.path.dirname(os.path.abspath(str(path))), + os.path.basename(str(path)), + dir_key=dir_key, + ) + + +def nz_config_stanza( + directory, filename, *, dir_key="zlens_directory", value_loc="mid" +): + """Build the OneCovariance ``[redshift]`` config stanza for an n(z) file. + + Parameters + ---------- + directory : str + Directory holding the n(z) file (OneCovariance ``*_directory`` value). + filename : str + n(z) file basename (OneCovariance ``zlens_file`` value). + dir_key : str, optional + Directory config key — ``"zlens_directory"`` (upstream) or + ``"z_directory"`` (UNIONS template). Default ``"zlens_directory"``. + value_loc : str, optional + ``value_loc_in_lensbin`` — where in each histogram bin the tabulated + ``n(z)`` value sits (``mid``/``left``/``right``). Default ``"mid"``, + matching the bin-centred grids the SACC stores. + + Returns + ------- + dict + The ``[redshift]`` key/value pairs: ``{dir_key: directory, "zlens_file": + filename, "value_loc_in_lensbin": value_loc}``. Assign these under + ``config["redshift"]`` of a OneCovariance ``configparser`` config. + """ + if value_loc not in ("mid", "left", "right"): + raise ValueError( + f"value_loc_in_lensbin must be 'mid', 'left' or 'right'; got {value_loc!r}" + ) + return { + dir_key: directory, + "zlens_file": filename, + "value_loc_in_lensbin": value_loc, + } + + +def read_nz(path): + """Read a OneCovariance combined n(z) file back to ``(z, nz_columns)``. + + Inverse of :func:`write_nz` (the numeric round-trip; the config stanza is + not stored in the file). Comment/header lines are skipped. + + Parameters + ---------- + path : str or pathlib.Path + n(z) text file (column 0 = z, columns 1… = per-bin n(z)). + + Returns + ------- + tuple + ``(z, nz)`` where ``z`` is the shared redshift grid (shape ``(n_z,)``) + and ``nz`` is the per-bin distributions (shape ``(n_z, n_bins)``). + """ + table = np.atleast_2d(np.genfromtxt(str(path))) + return table[:, 0], table[:, 1:] + + +def covariance_blocks(cov_list, selectors, *, gaussian=True): + """Reshape a OneCovariance ``covariance_list`` table into SACC cov blocks. + + OneCovariance emits a flat ``covariance_list_*.dat`` table with one row per + ``(i, j)`` element pair (row-major, ``k = i·n + j``); the covariance value + lives in column 10 (Gaussian) or column 9 (Gaussian+non-Gaussian). This + reshapes the flat table into dense square block(s) — reusing + :func:`sp_validation.statistics.cov_from_one_covariance` for the per-block + reshape — and pairs each with its SACC selector, ready for + :func:`sp_validation.sacc_io.assemble_covariance`. + + Single-statistic case: pass the whole table and one selector; you get one + ``(selector, dense)`` block. Multi-statistic case (tomography-ready): pass a + sequence of ``(selector, sub_table)`` pairs — each ``sub_table`` a + contiguous slice of the flat output for one statistic / bin-pair — and each + is reshaped and re-paired with its selector in order. The API is thus shaped + to extend to multi-probe blocking without over-fitting the single-bin case. + + Parameters + ---------- + cov_list : numpy.ndarray or sequence + Either the flat OneCovariance table (2-D array, one row per pair) for a + single block, or — for the multi-block form — a sequence of + ``(selector, sub_table)`` pairs. In the multi-block form ``selectors`` + must be ``None`` (the selectors travel with the sub-tables). + selectors : selector or None + For the single-block form, the SACC selector for the whole table (a + ``(data_type, tracers[, tags])`` tuple or an index array, as + :func:`sacc_io.assemble_covariance` accepts). Must be ``None`` for the + multi-block form. + gaussian : bool, optional + Select the Gaussian-only column (``True``, default) or the + Gaussian+non-Gaussian column (``False``); passed straight through to + ``cov_from_one_covariance``. + + Returns + ------- + list + Ordered ``(selector, dense_cov)`` pairs, directly consumable by + ``sacc_io.assemble_covariance(s, blocks)``. + """ + if selectors is None: + # Multi-block form: cov_list is a sequence of (selector, sub_table). + return [ + (selector, cov_from_one_covariance(np.asarray(sub), gaussian=gaussian)) + for selector, sub in cov_list + ] + # Single-block form: one flat table, one selector. + return [ + (selectors, cov_from_one_covariance(np.asarray(cov_list), gaussian=gaussian)) + ] diff --git a/src/sp_validation/tests/test_one_covariance_io.py b/src/sp_validation/tests/test_one_covariance_io.py new file mode 100644 index 00000000..a99ad552 --- /dev/null +++ b/src/sp_validation/tests/test_one_covariance_io.py @@ -0,0 +1,274 @@ +"""Tests for :mod:`sp_validation.one_covariance_io`. + +All synthetic, all fast: the OneCovariance fixtures are built in memory +shaped exactly like its real file I/O — a flat ``covariance_list`` table with +the ``(i, j)`` index rows and the Gaussian / Gauss+non-Gaussian value columns +at the indices ``cov_from_one_covariance`` expects (col 10 / col 9), and the +combined n(z) text file (column 0 = z, one column per bin, no edges). No +cluster paths; OneCovariance is not imported. + +The two pieces: + +- **Piece 1** (n(z) SACC -> OneCovariance input): write the combined n(z) + file from a SACC's ``source_i`` NZ tracers, read it back, assert the z and + n(z) columns round-trip and the config stanza names the file. +- **Piece 2** (OneCovariance output -> SACC covariance blocks): reshape the + flat table to the dense block(s) and prove they feed + ``sacc_io.assemble_covariance`` cleanly (reshaped block -> assemble -> + ``s.covariance.dense`` matches the hand-built matrix). +""" + +import numpy as np +import numpy.testing as npt +import pytest + +from sp_validation import one_covariance_io as ocio +from sp_validation import sacc_io as sio + + +# --------------------------------------------------------------------------- # +# Synthetic OneCovariance-shaped fixtures +# --------------------------------------------------------------------------- # +def _one_cov_table(cov_gauss, cov_all): + """Flatten two n x n matrices into a OneCovariance ``covariance_list`` table. + + Reproduces the real flat output: one row per ``(i, j)`` element pair in + row-major order ``k = i·n + j``, with the Gaussian value in column 10 and + the Gaussian+non-Gaussian value in column 9. Columns 0-8 and the index + columns are filled with self-documenting placeholder values (the reshape + only reads cols 9/10, but a realistic width proves it does not spill). + """ + n = cov_gauss.shape[0] + rows = [] + for i in range(n): + for j in range(n): + row = np.arange(11.0) # placeholder cols 0-8 (+ overwritten 9,10) + row[9] = cov_all[i, j] + row[10] = cov_gauss[i, j] + rows.append(row) + return np.array(rows) + + +def _spd(n, seed): + """Symmetric positive-definite matrix of size ``n`` (a valid covariance).""" + a = np.random.default_rng(seed).normal(size=(n, n)) + return a @ a.T + n * np.eye(n) + + +def _nz(seed, n=40): + rng = np.random.default_rng(seed) + z = np.linspace(0.01, 2.0, n) + return z, rng.uniform(0.1, 1.0, n) + + +# --------------------------------------------------------------------------- # +# Piece 2 — flat covariance_list -> dense block (reshape correctness + teeth) +# --------------------------------------------------------------------------- # +def test_covariance_blocks_reshapes_to_hand_built_matrix(): + """Pin the reshape and prove gaussian vs gauss+ng select different columns. + + WHAT IS PINNED: ``covariance_blocks`` flattens/reshapes the OneCovariance + ``covariance_list`` table into a dense square block matching a hand-built + covariance. It delegates the per-block reshape to + ``statistics.cov_from_one_covariance``, so column 10 (gaussian) and column 9 + (gauss+ng) must recover the two distinct hand-built matrices. + + WHY TEETH: (a) ``gaussian=True`` vs ``False`` must return the two *different* + matrices, proving the column flag is load-bearing; (b) perturbing a single + entry of the flat input must change exactly that entry of the reshaped + block, proving the reshape actually reads the table (not a constant). + """ + cov_gauss = _spd(4, seed=1) + cov_all = _spd(4, seed=2) + table = _one_cov_table(cov_gauss, cov_all) + + selector = (sio.XI_PLUS, (sio.source_name(0), sio.source_name(0))) + + [(sel_g, block_g)] = ocio.covariance_blocks(table, selector, gaussian=True) + [(sel_a, block_a)] = ocio.covariance_blocks(table, selector, gaussian=False) + + assert sel_g == selector and sel_a == selector + npt.assert_allclose(block_g, cov_gauss, rtol=1e-12) + npt.assert_allclose(block_a, cov_all, rtol=1e-12) + + # TEETH: gaussian and gauss+ng select different columns -> different blocks. + assert not np.allclose(block_g, block_a) + + # TEETH: a perturbation of one flat-table entry moves exactly that block + # entry (row k = i·n + j, col 10 for gaussian). + perturbed = table.copy() + perturbed[2 * 4 + 1, 10] += 5.0 # element (i=2, j=1) + [(_, block_p)] = ocio.covariance_blocks(perturbed, selector, gaussian=True) + npt.assert_allclose(block_p[2, 1] - block_g[2, 1], 5.0, rtol=1e-12) + block_p[2, 1] = block_g[2, 1] + npt.assert_allclose(block_p, block_g, rtol=1e-12) # nothing else moved + + +def test_covariance_blocks_multiblock_form(): + """Prove the tomography-ready multi-block form reshapes each sub-table. + + WHAT IS PINNED: passing ``selectors=None`` and a sequence of + ``(selector, sub_table)`` pairs reshapes each sub-table independently and + returns them paired with their selectors in order — the shape needed to map + a multi-statistic OneCovariance output onto several SACC selectors. + + WHY TEETH: the two sub-tables carry different matrices; if the function + reshaped only the first or mixed them, the second block would not match its + own hand-built matrix. + """ + cov_a, cov_b = _spd(3, seed=3), _spd(2, seed=4) + table_a = _one_cov_table(cov_a, _spd(3, seed=5)) + table_b = _one_cov_table(cov_b, _spd(2, seed=6)) + sel_a = (sio.XI_PLUS, (sio.source_name(0), sio.source_name(0))) + sel_b = (sio.XI_MINUS, (sio.source_name(0), sio.source_name(0))) + + blocks = ocio.covariance_blocks( + [(sel_a, table_a), (sel_b, table_b)], None, gaussian=True + ) + + assert [s for s, _ in blocks] == [sel_a, sel_b] + npt.assert_allclose(blocks[0][1], cov_a, rtol=1e-12) + npt.assert_allclose(blocks[1][1], cov_b, rtol=1e-12) + + +# --------------------------------------------------------------------------- # +# Piece 2 — the reshaped block feeds assemble_covariance cleanly +# --------------------------------------------------------------------------- # +def test_covariance_blocks_feed_assemble_covariance(): + """Round-trip: reshaped block -> assemble_covariance -> dense matches. + + WHAT IS PINNED: the ``(selector, dense)`` pair ``covariance_blocks`` + returns is directly consumable by ``sacc_io.assemble_covariance``: assembled + onto a SACC whose only statistic is one ξ+ block, ``s.covariance.dense`` + must equal the hand-built OneCovariance matrix. This is the end-to-end + contract between the two modules. + + WHY TEETH: the block must tile the data vector exactly; if the reshape + produced the wrong size or the wrong selector, ``assemble_covariance`` would + raise (its contiguity/tiling/size checks), so a clean assemble + matching + dense is a real proof. + """ + theta = np.geomspace(1.0, 100.0, 4) + xip, xim = np.arange(4) * 1e-5, np.arange(4) * 2e-5 + s = sio.new_sacc({0: _nz(0)}) + sio.add_xi(s, (0, 0), theta, xip, xim, grid="coarse") + + # The ξ block spans ξ+ then ξ− for the pair -> 8 points, one contiguous + # block (pair-major, matching sacc_io's canonical order). + cov_gauss = _spd(8, seed=7) + table = _one_cov_table(cov_gauss, _spd(8, seed=8)) + pair = (sio.source_name(0), sio.source_name(0)) + selector = np.concatenate( + [s.indices(sio.XI_PLUS, pair), s.indices(sio.XI_MINUS, pair)] + ) + + blocks = ocio.covariance_blocks(table, selector, gaussian=True) + sio.assemble_covariance(s, blocks) + + npt.assert_allclose(s.covariance.dense, cov_gauss, rtol=1e-12) + + +# --------------------------------------------------------------------------- # +# Piece 1 — n(z) SACC -> OneCovariance input (round-trip + config stanza) +# --------------------------------------------------------------------------- # +def test_write_nz_roundtrips_and_names_file(tmp_path): + """Round-trip the n(z) file and check the config stanza names it. + + WHAT IS PINNED: ``write_nz`` writes the SACC ``source_i`` NZ tracers as the + OneCovariance combined file (column 0 = z, one column per bin, no z_low/ + z_high edges). ``read_nz`` recovers the z grid and every per-bin n(z) + column, and the returned ``[redshift]`` stanza names the exact file written + (directory + basename) plus ``value_loc_in_lensbin``. + + WHY TEETH: the z grid and each n(z) column must round-trip to the values the + SACC holds (drawn from a seeded RNG); a transposed write or a dropped column + would fail the per-bin comparison. The stanza's directory/file must match + the path actually written. + """ + z, nz0 = _nz(10) + _, nz1 = _nz(11) + s = sio.new_sacc({0: (z, nz0), 1: (z, nz1)}) + + path = tmp_path / "nz_onecov.txt" + stanza = ocio.write_nz(s, path, n_bins=2) + + z_read, nz_read = ocio.read_nz(path) + npt.assert_allclose(z_read, z, rtol=1e-12) + assert nz_read.shape == (len(z), 2) + npt.assert_allclose(nz_read[:, 0], nz0, rtol=1e-12) + npt.assert_allclose(nz_read[:, 1], nz1, rtol=1e-12) + + assert stanza["zlens_directory"] == str(tmp_path) + assert stanza["zlens_file"] == "nz_onecov.txt" + assert stanza["value_loc_in_lensbin"] == "mid" + + +def test_write_nz_unions_template_dir_key(tmp_path): + """The UNIONS-template ``z_directory`` key is selectable via ``dir_key``. + + WHAT IS PINNED: the upstream OneCovariance key is ``zlens_directory``, but + the UNIONS template (pseudo_cl.py._modify_onecov_config) writes + ``z_directory``. ``dir_key="z_directory"`` produces that variant so the + stanza drops straight into the UNIONS template's ``[redshift]`` section. + """ + z, nz0 = _nz(12) + s = sio.new_sacc({0: (z, nz0)}) + stanza = ocio.write_nz(s, tmp_path / "nz.txt", n_bins=1, dir_key="z_directory") + assert "z_directory" in stanza and "zlens_directory" not in stanza + assert stanza["z_directory"] == str(tmp_path) + assert stanza["zlens_file"] == "nz.txt" + + +def test_write_nz_fails_on_mismatched_z_grids(tmp_path): + """Fail fast when source bins do not share one redshift grid. + + WHAT IS PINNED: the OneCovariance combined file has a single redshift + column, so all bins must share the z grid. A bin on a different grid must + raise ``ValueError`` at write time, not silently mis-align. + """ + z0, nz0 = _nz(20) + z1_shifted, nz1 = _nz(21) + z1_shifted = z1_shifted + 0.1 # different grid + s = sio.new_sacc({0: (z0, nz0), 1: (z1_shifted, nz1)}) + with pytest.raises(ValueError, match="differs from source bin 0"): + ocio.write_nz(s, tmp_path / "bad.txt", n_bins=2) + + +def test_write_nz_no_header_roundtrips(tmp_path): + """The bare (header=False) file still round-trips numerically. + + WHAT IS PINNED: ``header=False`` writes a purely numeric file (no ``#`` + column-name line); ``read_nz`` recovers the same z grid and n(z) column, so + the header is cosmetic and never load-bearing for the numeric round-trip. + """ + z, nz0 = _nz(40) + s = sio.new_sacc({0: (z, nz0)}) + path = tmp_path / "bare.txt" + ocio.write_nz(s, path, n_bins=1, header=False) + z_read, nz_read = ocio.read_nz(path) + npt.assert_allclose(z_read, z, rtol=1e-12) + npt.assert_allclose(nz_read[:, 0], nz0, rtol=1e-12) + + +def test_nz_config_stanza_rejects_bad_value_loc(): + """``value_loc_in_lensbin`` outside {mid,left,right} fails fast. + + WHAT IS PINNED: OneCovariance only accepts ``mid``/``left``/``right`` for + the histogram-bin value location; an invalid value is a config bug and must + raise ``ValueError`` rather than write a stanza OneCovariance will reject. + """ + with pytest.raises(ValueError, match="value_loc_in_lensbin"): + ocio.nz_config_stanza("/dir", "nz.txt", value_loc="center") + + +def test_write_nz_fails_on_missing_bin(tmp_path): + """Fail fast when a requested source bin is absent from the SACC. + + WHAT IS PINNED: requesting more bins than the SACC carries is a real config + bug; ``write_nz`` raises ``ValueError`` naming the missing tracer rather + than writing a short file. + """ + z, nz0 = _nz(30) + s = sio.new_sacc({0: (z, nz0)}) + with pytest.raises(ValueError, match="source_1"): + ocio.write_nz(s, tmp_path / "short.txt", n_bins=2) From 700a88579942ac5afb6574c1949d6fa2a55fb85a Mon Sep 17 00:00:00 2001 From: Cail Daley Date: Fri, 10 Jul 2026 07:34:24 +0200 Subject: [PATCH 3/3] =?UTF-8?q?fix(twopoint):=20fail=20fast=20on=20tomogra?= =?UTF-8?q?phic/=CE=BE-less=20SACC;=20pin=20covariance=20gather?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Adversarial-review hardening (one HIGH, one MEDIUM, one coverage gap): - HIGH: n_bins>1 silently truncated — n_bins drove only the NZDATA column count while data/covariance were read from bin (0, 0), so a 2-bin SACC emitted a plausible-looking FITS carrying 1/3 of the data. The converter now fails fast unless n_bins == 1 and the ξ tracer pairs are exactly {(source_0, source_0)}; tomographic emission lands with the tomographic round. - MEDIUM: a ξ-less SACC wrote an empty XI_PLUS and a (0, 0) COVMAT silently; now a loud ValueError. - Coverage: every prior covariance test exercised the identity permutation (single-pair SACC order is already type-major). A (row, col)-encoded covariance test now pins the exact np.ix_ gather of COVMAT (block-diag ξ + joint non-adjacent [τ0+; τ2+]) and COVMAT_CELL — any transposition, offset, or swapped block fails. Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_016Ko89uZF84ez6HEDcTWbJe --- .../tests/test_twopoint_convert.py | 94 +++++++++++++++++++ src/sp_validation/twopoint_convert.py | 42 ++++++++- 2 files changed, 134 insertions(+), 2 deletions(-) diff --git a/src/sp_validation/tests/test_twopoint_convert.py b/src/sp_validation/tests/test_twopoint_convert.py index 5e599717..3a5bc44a 100644 --- a/src/sp_validation/tests/test_twopoint_convert.py +++ b/src/sp_validation/tests/test_twopoint_convert.py @@ -376,3 +376,97 @@ def test_rho_tau_sidecars_required_together(tmp_path): twopoint_convert.sacc_to_twopoint_fits( s, str(tmp_path / "x.fits"), rho_stats_hdu=rho_hdu, n_bins=1 ) + + +# ============================================================================= +# Fail-fast guards and permutation teeth (adversarial-review hardening) +# ============================================================================= + + +def test_tomographic_sacc_raises(tmp_path): + """A multi-bin SACC fails fast instead of silently truncating to (0, 0). + + Review finding (HIGH): ``n_bins`` alone drove the NZDATA column count while + the data vector and covariance were read from bin ``(0, 0)`` only, so a + 2-bin SACC + ``n_bins=2`` emitted a plausible-looking FITS carrying 1/3 of + the data. Both the ``n_bins`` and the tracer-pair mismatch must raise. + """ + inp = _inputs(seed=30) + s = sacc_io.new_sacc({0: (inp["z"], inp["nz"]), 1: (inp["z"], inp["nz"])}) + for pair in [(0, 0), (0, 1), (1, 1)]: + sacc_io.add_xi(s, pair, inp["theta"], inp["xip"], inp["xim"], grid="coarse") + s.add_covariance(np.eye(len(s.mean))) + + with pytest.raises(ValueError, match="single-bin only"): + twopoint_convert.sacc_to_twopoint_fits(s, str(tmp_path / "x.fits"), n_bins=2) + with pytest.raises(ValueError, match="single-bin only"): + twopoint_convert.sacc_to_twopoint_fits(s, str(tmp_path / "x.fits"), n_bins=1) + assert not (tmp_path / "x.fits").exists() + + +def test_sacc_without_xi_raises(tmp_path): + """A SACC with no ξ± points raises instead of writing an empty data vector.""" + inp = _inputs(seed=31) + s = sacc_io.new_sacc({0: (inp["z"], inp["nz"])}) + sacc_io.add_pseudo_cl( + s, + (0, 0), + inp["ell"], + inp["cl_ee"], + inp["cl_bb"], + inp["cl_eb"], + window_ells=np.arange(2, 102), + window_weights=np.random.default_rng(9).uniform(0, 1, (100, N_ELL)), + ) + s.add_covariance(np.eye(len(s.mean))) + + with pytest.raises(ValueError, match="nothing to convert"): + twopoint_convert.sacc_to_twopoint_fits(s, str(tmp_path / "x.fits")) + assert not (tmp_path / "x.fits").exists() + + +def test_covmat_blocks_exact_gather_encoded_cov(tmp_path): + """Every COVMAT/COVMAT_CELL entry is the exact ``np.ix_`` gather of the SACC + covariance, pinned with a (row, col)-encoded matrix. + + Review finding (MEDIUM): for a single bin pair the ξ gather happens to be + the identity permutation, so the byte-compares alone could pass with a + transposed or block-swapped gather. Encoding ``C[i, j] = i*n + j`` (asymmetric, + every entry unique) makes any transposition, offset, or wrong block produce + detectably wrong values; the τ gather is genuinely non-identity (τ_0− sits + between τ_0+ and τ_2+ in insertion order). Expected layout per + ``covdat_to_fits``: block_diag(ξ type-major gather, joint [τ_0+; τ_2+] + gather), with COVMAT_CELL the CELL_EE gather in its own HDU. + """ + inp = _inputs(seed=32) + s = _sacc(inp, cl=True, rho_tau=True) + n = len(s.mean) + encoded = np.arange(n * n, dtype=float).reshape(n, n) + s.add_covariance(encoded, overwrite=True) + rho_hdu, tau_hdu = _sidecar_hdus(tmp_path, inp) + + out = tmp_path / "encoded.fits" + twopoint_convert.sacc_to_twopoint_fits( + s, str(out), rho_stats_hdu=rho_hdu, tau_stats_hdu=tau_hdu, n_bins=1 + ) + + pair = (SOURCE, SOURCE) + xi_idx = np.concatenate( + [s.indices(sacc_io.XI_PLUS, pair), s.indices(sacc_io.XI_MINUS, pair)] + ) + tau_idx = np.concatenate( + [ + s.indices(sacc_io.TAU_PLUS.format(k=0), (SOURCE, PSF)), + s.indices(sacc_io.TAU_PLUS.format(k=2), (SOURCE, PSF)), + ] + ) + expected = twopoint_convert._block_diag( + encoded[np.ix_(xi_idx, xi_idx)], encoded[np.ix_(tau_idx, tau_idx)] + ) + cell_idx = s.indices(sacc_io.CL_EE, pair) + + with fits.open(out) as hdul: + np.testing.assert_array_equal(hdul["COVMAT"].data, expected) + np.testing.assert_array_equal( + hdul["COVMAT_CELL"].data, encoded[np.ix_(cell_idx, cell_idx)] + ) diff --git a/src/sp_validation/twopoint_convert.py b/src/sp_validation/twopoint_convert.py index c06e0eb3..8f787f3b 100644 --- a/src/sp_validation/twopoint_convert.py +++ b/src/sp_validation/twopoint_convert.py @@ -160,6 +160,33 @@ def _type_major_xi(s, bins): return sacc_io.get_xi(s, bins, grid="coarse") +def _require_single_bin(s, n_bins): + """Fail fast unless the SACC is a valid single-bin ξ product. + + The converter emits the single-bin 2pt-FITS today's CosmoSIS pipeline reads + (BIN1/BIN2 all 1, one NZ column). A tomographic SACC would otherwise slip + through silently — ``n_bins`` alone drives the NZDATA column count while the + ξ/covariance are read from bin ``(0, 0)`` only, so a 2-bin file would emit a + ``NBIN=2`` n(z) beside a data vector holding just the ``(0, 0)`` pair. + Guards both the empty-ξ case and the single-bin contract; tomographic + emission lands with the tomographic round. + """ + pairs = s.get_tracer_combinations(sacc_io.XI_PLUS) + if not pairs: + raise ValueError( + f"SACC has no {sacc_io.XI_PLUS} points — nothing to convert; the " + "2pt-FITS data vector is built from the ξ± statistics" + ) + expected = (sacc_io.source_name(0), sacc_io.source_name(0)) + if n_bins != 1 or set(pairs) != {expected}: + raise ValueError( + f"converter is single-bin only (n_bins=1, ξ pairs == {{{expected}}}); " + f"got n_bins={n_bins} and ξ pairs {sorted(pairs)}. Tomographic " + "emission (multiple bin pairs, per-pair BIN1/BIN2, one NZ column per " + "bin) lands with the tomographic round." + ) + + def sacc_to_twopoint_fits( s, path, @@ -191,19 +218,30 @@ def sacc_to_twopoint_fits( alone cannot rebuild the ``varrho_*`` columns Sacha's fork reads. When omitted, a pure ξ (± Cℓ) product is written. n_bins : int, optional - Number of source tomographic bins (default 1, the current single-bin - analysis). Sets the NZDATA column count. + Number of source tomographic bins. Must be ``1``: this converter emits + the single-bin 2pt-FITS today's CosmoSIS pipeline consumes. Tomographic + emission (multiple bin pairs, per-pair BIN1/BIN2, one NZ column per bin) + lands with the tomographic round; the converter fails fast on anything + else rather than silently truncating to bin ``(0, 0)``. Returns ------- astropy.io.fits.HDUList The assembled list, also written to ``path``. + + Raises + ------ + ValueError + If the SACC has no ξ points; if ``n_bins != 1`` or the SACC's ξ tracer + pairs are anything other than exactly ``{(source_0, source_0)}`` (the + single-bin contract); or if exactly one of the ρ/τ sidecars is supplied. """ if (rho_stats_hdu is None) != (tau_stats_hdu is None): raise ValueError( "rho_stats_hdu and tau_stats_hdu must be supplied together " "(the ρ/τ product needs both, or neither for a pure-ξ product)" ) + _require_single_bin(s, n_bins) use_rho_tau = rho_stats_hdu is not None bins = (0, 0)