{"id":"679c2e0f-4db6-4296-863e-4ef75f06834a","arxiv_id":"2607.23494","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Hermes/PyHermes reconstructs catalogues in a scaling-function basis and unifies CIC, 2PCF, 3PCF, marked, and operator-based cosmic statistics as reusable window operations with FFT/MPI/GPU scaling.","lead":"Hermes turns galaxy catalogues into multiresolution density fields and measures clustering by applying window filters instead of counting pairs and triplets. That design lets one reconstructed field feed many statistics and scales better for the huge catalogues coming from next-generation surveys.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The reuse/scalability claim rests on translation-invariant FFT convolution demonstrated only in periodic boxes with analytic uniform randoms; nothing in §4 shows the estimator identities or the O(N_grid log N_grid) cost survive realistic masks and survey geometry.","rationale":"Read in good faith, this is a careful methods paper with genuine independent support: cross-validation of the isotropic 2PCF against pycorr/Corrfunc and TreeCorr (Fig. 13), matter PDFs against pyLDT-cosmo (Fig. 6), angular-3PCF/multipole consistency checks (Fig. 11), emulator cross-checks on the Kun suite, publicly released code, and honest reporting of the finite-J small-scale suppression and the J→J+1 memory factor of eight. The reader's CONDITIONAL verdict with HIGH confidence and low correctness risk is well calibrated. The single most load-bearing concern is exactly the one the reader flagged: the framework's computational identities (Toeplitz FFT convolution, difference-field LS estimator, analytic uniform randoms) are demonstrated only under periodic boundary conditions without masks, while the paper's stated purpose is application to real survey catalogues. The concern is not that the formalism is wrong — the kernel's translation invariance is a property of the window, not the field, so masked random fields can in principle be projected and convolved — but that periodic wraparound, bounding-box overhead, and finite-J accuracy of the random-field RR term are unquantified, and the paper's own claim that geometry \"can be incorporated within the same framework\" is asserted, not shown. Because the reader's weakest_assumption already names this gap and priced it into CONDITIONAL, my read does not move the verdict; the masked-mock test above is the concrete condition that should govern upgrading from CONDITIONAL to ACCEPT.","tokens_in":59034,"tokens_out":3283,"duration_ms":138325,"concrete_test":"Cut a Quijote halo catalogue to a realistic geometry: a spherical-cap or wedge angular mask plus a radial selection function n(z), with ~20× randoms drawn in the same masked volume. Embed data and randoms in a zero-padded box (L→2L, i.e. one extra dilation level) to suppress FFT wraparound, project both into the MRA basis, and measure the LS 2PCF with PyHermes using the projected random field for DR and RR. Compare against Corrfunc on the identical masked catalogues. If PyHermes matches Corrfunc to ≲1% above a few cell widths, including bins near the mask boundary, and the padded run stays within the advertised memory envelope, the survey-geometry transfer holds; systematic edge offsets, or failure unless padding is used, would show the periodic-FFT architecture does not transfer at the claimed cost.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim — one MRA reconstruction reused for many statistics, cost set by grid resolution and window count, \"well suited to large data sets from current and future galaxy surveys\" — is carried by two ingredients: (i) the Toeplitz structure of the kernel matrix (Eq. 60, w^j_{l,m}=Φ_{l−m}(r)) that licenses FFT convolution (Eqs. 62, 110), and (ii) the difference-field Landy–Szalay trick (Eqs. 55–57) with randoms handled analytically (task.random=\"uniform\"). Both are validated only in periodic 1000 h^-1 Mpc boxes; §4 states plainly that \"all demonstrations assume periodic boundary conditions and do not include survey masks, angular selection functions, or spatially varying random catalogues.\" Two specific gaps follow. First, periodic FFT convolution wraps the domain: for a non-periodic survey footprint the kernel is still translation-invariant, but filtered fields near the footprint edge pick up wrapped contributions unless the field is embedded in a zero-padded bounding box — up to 8× more cells in 3D, and for a cone-like survey the cost is then set by the bounding-box volume, not the surveyed volume. This directly erodes the memory/cost advantage that is the paper's raison d'être (Table 3 already shows 61.5 GiB host memory at J=9 in a 1000^3 box without padding). Second, with a masked random catalogue the RR[W] denominator of Eq. 57 must itself be computed from a projected random field, whose finite-J accuracy at small separations is an unvalidated error source. The paper asserts realistic geometry \"can be incorporated within the same framework\" but provides no masked-mock demonstration, so transfer of both the estimator identities and the headline scaling is assumed rather than shown. The finite-J small-scale suppression (Fig. 13) is real but quantified and convergent; it is not the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":null,"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a usable computational-methods paper, not a cosmology result. They take MRACS (Feng 2007), the in-situ filtered-field view (Yue et al. 2024), and the multipole 3PCF work (Ju et al. 2026), cast a broad class of LSS stats as window operators on one multiresolution reconstruction, and ship PyHermes with MPI/threads/GPU and public code.\n\nWhat is actually new is the packaging and the breadth: one SFC field reused for CIC, isotropic/anisotropic 2PCF, marked and wavelet stats, standard and multipole 3PCF, plus differential/potential operators, with systematic timing/memory numbers. That is real engineering value for mock suites and multi-statistic pipelines.\n\nWhat they do well: external checks. 2PCF converges to pycorr/Corrfunc/TreeCorr as J rises (Fig. 13). CIC PDFs match pyLDT. Angular vs multipole 3PCF consistency is shown. Marked/wavelet/3PCF monopoles track Kun emulators as comparison targets, not circular proofs. Operator windows recover the expected large-scale continuity/Poisson patterns. The math (Toeplitz kernels → FFT convolution, difference-field LS, connection-coefficient contractions) is coherent and cited to prior work honestly.\n\nSoft spots, in proportion. Finite-J small-scale suppression is real and they show it; it is a resolution choice, not a hidden bug. The stress-test point lands but is not fatal: all body demos are periodic boxes with analytic uniform randoms; masks, selection, and varying randoms are asserted as “same framework” without a masked-mock figure. Padding a survey footprint will inflate N_grid and can eat the memory advantage (Table 3 already ~60 GiB at J=9). That limits the “future surveys” claim until someone runs it; it does not invalidate the periodic-box identities or the code as shipped. Title’s “optimal” is marketing; the content is “unified and scalable under stated assumptions.”\n\nWho it is for: people building multi-probe LSS pipelines, mock factories, or custom windows (marks, wavelets, operators). Not required reading if you only need a standard pair counter on a small catalogue.\n\nI would send this to peer review. Ask referees for a masked/random-catalogue demo or a clear cost model with padding, and dial back “optimal,” but the work is serious and citeable as methods infrastructure.","headline":"Solid methods/software paper: unified window-on-MRA-field language plus a real PyHermes stack, externally cross-checked; survey-geometry demos are missing but the core math holds in the regime they actually test.","tokens_in":55230,"tokens_out":642,"would_cite":true,"duration_ms":20601,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"One reconstructed density field plus window kernels replaces separate tuple-counting algorithms for cosmic clustering statistics.","keywords":["large-scale structure","correlation functions","multiresolution analysis","window functions","counts-in-cells","three-point correlation function","N-body catalogues","high-performance computing"],"falsifier":"Measure the same isotropic 2PCF and multipole 3PCF on a large halo catalogue with Hermes at high resolution and with a standard pair/triplet counter on identical bins; if small-scale amplitudes or multipoles disagree beyond resolution-controlled residuals, or if adding a realistic mask and random catalogue breaks the field–window reuse without new cost blow-ups, the central practical claim fails.","tokens_in":54886,"feed_emoji":"🌌","tokens_out":895,"duration_ms":28646,"temperature":0.7,"pith_summary":"Hermes claims that cosmic large-scale-structure statistics need not each be a separate particle-counting program. A galaxy or halo catalogue is projected once onto a continuous multiresolution density field built from compact scaling functions. Counts-in-cells, two-point and three-point correlations, multipoles, marked statistics, and even derived fields such as potential and velocity divergence are then obtained by convolving that field with chosen window kernels and forming products of the filtered fields. Because the expensive step is the field reconstruction and FFT convolutions on a grid, cost tracks grid resolution and the number of windows rather than the combinatorial explosion of N-tuples. A single field can therefore be reused for many standard and custom measurements, which matters for the huge catalogues expected from current and future galaxy surveys.","feed_headline":"One density field, many cosmic statistics via windows","feed_subtitle":"Hermes turns pair and triplet counts into reusable filtered-field products for huge galaxy catalogues","key_machinery":"The field–window architecture: catalogue → multiresolution scaling-function coefficients → FFT convolution with a chosen window (shell, ring, multipole, mark, derivative, inverse Laplacian, etc.) → inner products of filtered fields. Different statistics are different kernels, not different counting codes.","core_discovery":"The paper establishes that conventional cosmic statistics are special cases of window operations on one multiresolution density field: reconstructing a discrete catalogue in a compact scaling-function basis and replacing explicit pair/triplet counting with algebraic products of window-filtered fields yields a unified, reusable estimator whose computational cost is controlled by field resolution and window count rather than by the number of particle tuples.","pith_inferences":["The same separation of representation from measurement could extend cleanly to four-point and density-split pipelines if higher-fold field products stay memory-feasible.","Survey-realistic randoms and masks will likely force either masked-field reconstructions or hybrid particle corrections; how much of the reuse advantage survives that step is an open stress test.","Fourier differential and inverse-Laplacian windows suggest a path to consistent density–velocity–potential diagnostics inside one codebase, useful for RSD and environment studies beyond pure clustering."],"forward_implications":["One catalogue projection can feed CIC PDFs, isotropic and redshift-space 2PCF, marked 2PCF, standard and multipole 3PCF, and operator-built potential/acceleration fields without rewriting estimators.","New statistics can be introduced by designing or composing window kernels rather than by inventing new N-tuple counters.","Runtime for high-order and multi-probe analyses of large surveys scales primarily with grid size and number of windows, not with N_g^N tuple enumeration.","Emulator and covariance pipelines that need many statistics on many mocks become cheaper once fields are cached and reused."],"fun_headline_variants":["One multiresolution field yields many cosmic statistics via windows","Hermes turns pair counts into reusable filtered-field algebra","Window kernels unify CIC, 2PCF and 3PCF on one density field","Reuse one catalogue reconstruction for standard and custom stats","Field resolution, not particle tuples, sets the compute cost"],"cache_read_input_tokens":49280,"weakest_assumption_plain":"A finite-resolution multiresolution projection of a periodic box is taken as good enough to stand in for the true continuous field and for real survey geometry, even though the paper’s demos omit masks, selection functions, and varying random catalogues and show small-scale correlation loss at low resolution.","fun_headline_variants_meta":{"raw":{"variants":["One multiresolution field yields many cosmic statistics via windows","Hermes turns pair counts into reusable filtered-field algebra","Window kernels unify CIC, 2PCF and 3PCF on one density field","Reuse one catalogue reconstruction for standard and custom stats","Field resolution, not particle tuples, sets the compute cost"]},"model":"grok-4.5","effort":"low","cost_usd":0.002213,"raw_usage":{"total_tokens":944,"prompt_tokens":754,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":22128000,"prompt_tokens_details":{"text_tokens":754,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":121,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":754,"tokens_out":69,"duration_ms":4470,"temperature":1.0,"reasoning_tokens":121,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T20:50:07.106484+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Measure the same isotropic 2PCF and multipole 3PCF on a large halo catalogue with Hermes at high resolution and with a standard pair/triplet counter on identical bins; if small-scale amplitudes or multipoles disagree beyond resolution-controlled residuals, or if adding a realistic mask and random catalogue breaks the field–window reuse without new cost blow-ups, the central practical claim fails.","supporting_citations":[],"review_version":1}