Large-scale structure (cosmology)
The distribution of galaxies carries information about dark matter, dark energy and the mass of neutrinos, and extracting it from survey data requires precise theory. This project takes that theory to the next loop order, two loops for galaxy clustering in redshift space and one loop for the non-Gaussian trispectrum, and presents KITE, a new AI-designed framework for fitting the cosmological parameters to data from DESI and other surveys.
Reading a galaxy map
A spectroscopic survey such as DESI records each galaxy’s direction and its redshiftthe stretching of a galaxy’s light by cosmic expansion; it grows with distance, so it serves as the third coordinate of the map, which gives its distance. Gravity has made the map lumpy, and cosmologists summarize the lumpiness with correlation statistics. The power spectrumthe two-point statistic: how much structure the map contains at each wavenumber k, where larger k means shorter wavelength $P(k)$ measures how much structure there is at each wavenumber $k$; the bispectrum does the same for triples of density ripples whose wavevectors close into a triangle, and the trispectrum for quadruples. All three depend on the composition of the universe, the growth of structure and the initial conditions.
Two features of the map complicate the model. First, galaxies form preferentially in dense regions, so the galaxy density is a biased tracera population whose clustering over- or under-represents the underlying matter; galaxies cluster more strongly than the dark matter they sit in of the matter density, related to it through free coefficients that stand in for the physics of galaxy formation. Second, a galaxy’s redshift includes its own motion on top of the cosmic expansion, so a galaxy falling toward a cluster is placed slightly too near or too far along the line of sight. Nick Kaiser worked out the leading effect of this redshift-space distortion in 1987; the distortion mixes the density field with the velocity field at every order of approximation.
A full-shape analysis fits a perturbative model, at present the effective field theory of large-scale structure at one-loop order, to the whole measured shape of these statistics. The work summarized here, by Matthew D. Schwartz, Mikhail M. Ivanov and Siddharth Mishra-Sharma, takes that model to the next order in two parts: the calculation itself (Paper I, posted on this site and marked preliminary) and KITE, the code that turns it into a likelihood for survey data (Paper II, in preparation).
The loop expansion
On the largest scales the density fluctuations are small and each Fourier mode grows independently. The power spectrum is then the linear spectrum $P_{\rm lin}$ that a Boltzmann codethe standard numerical solver that evolves the perturbations of the early universe and returns the linear power spectrum of a given cosmology such as CLASS computes for each cosmology. On small scales halos form and only simulations work; in between, the fluid equations for dark matter can be solved order by order in the initial fluctuation, by a recursion that Goroff, Grinstein, Rey and Wise wrote down in 1986. The $n$th-order density is an integral of a known kernel $F_n$ against $n$ copies of the linear field, and correlating two expanded fields gives the power spectrum as a series,
$$P(k) = P_{\rm lin}(k) + P_{\text{1-loop}}(k) + P_{\text{2-loop}}(k) + \cdots\,,\qquad P_{\text{2-loop}} = P_{15} + P_{24} + P_{33,\mathrm{I}} + P_{33,\mathrm{II}}\,.$$
The subscripts give the orders of the two fields being correlated: $P_{15}$ pairs the linear field with the fifth-order one, and the two $P_{33}$ terms are the two ways of pairing third-order fields. A “loop” is an integral over the wavenumber $\mathbf{q}$ of an intermediate mode that is summed over rather than observed, as in the loop diagrams of particle physics. The one-loop term is the sum of two such integrals,
$$P_{22}(k) = 2\!\int_{\mathbf{q}} \left[F_2(\mathbf{q},\mathbf{k}-\mathbf{q})\right]^2 P_{\rm lin}(q)\, P_{\rm lin}(|\mathbf{k}-\mathbf{q}|)\,,\qquad P_{13}(k) = 6\, P_{\rm lin}(k)\!\int_{\mathbf{q}} F_3(\mathbf{k},\mathbf{q},-\mathbf{q})\, P_{\rm lin}(q)\,,$$
where $\int_{\mathbf{q}}$ runs over all intermediate wavenumbers, $F_2$ and $F_3$ are the second- and third-order kernels, and the prefactors 2 and 6 count the ways of pairing the fields. Each further loop adds one linear spectrum and one more nested integral, so the two-loop term is smaller and much harder to compute.
The loop integrals run over all $\mathbf{q}$, including scales where dark matter is not a fluid at all. Around 2012 Baumann, Nicolis, Senatore and Zaldarriaga, and Carrasco, Hertzberg and Senatore, recast the expansion as an effective field theory, in which the unknown small-scale physics is absorbed into counterterms: extra terms of known shape in $k$ with coefficients fitted to the data. Galaxies enter through a bias expansion, a sum over every operator of the density and tidal fields that the symmetries allow, each with a free coefficient, following McDonald and Roy (2009) and later authors. Kaiser’s redshift-space map then turns the galaxy kernels into redshift-space kernels $Z_n^{\rm gal}$. Large-scale bulk flows smear the baryon acoustic feature in a way no fixed order captures, so that effect is handled by infrared resummationsumming the effect of large-scale bulk motions to all orders instead of order by order, which controls how strongly the acoustic feature is smeared, following Senatore and Zaldarriaga (2015) and Ivanov and Sibiryakov (2018). A stochastic shot-noise term accounts for galaxies being discrete objects. The one-loop model, with about ten free parameters per galaxy sample, was implemented in the public codes CLASS-PT and PyBird in 2020 and applied to the BOSS survey and then to DESI.
Why two loops now
In the one-loop analyses of BOSS and DESI the two-loop term appears only as an error estimate: the fitted range of wavenumbers is cut off so that the omitted term stays below the statistical error. One-loop fits to simulated galaxy samples with the cutoff pushed to higher wavenumbers (more aggressive scale cuts) return a shifted fluctuation amplitude, which may come from the missing two-loop term.
Most ingredients of the two-loop galaxy calculation existed for matter only or in real space only. Carrasco, Foreman, Green and Senatore computed the two-loop matter power spectrum in the effective theory in 2014, and Taule and Garny extended it to redshift space in 2023, still for matter. For biased tracers the two-loop power spectrum is known in real space (true positions, without the redshift displacement), where the bias expansion complete through fifth order has 29 independent operators, 17 of them with free coefficients at two loops. A joint analysis needs all three statistics for galaxies in redshift space on one operator basis, so that each bias coefficient means the same thing in all three. It also needs every loop integral evaluated to a known accuracy, because the two-loop term may itself be only a few percent of the signal.
Paper I: the calculation
The next order consists of three statistics of biased tracers in redshift space: the two-loop power spectrum, the one-loop bispectrum and the one-loop trispectrum. All three share one operator basis complete through fifth order and one set of redshift-space kernels. The bispectrum is renormalizedmoving the parts of a loop integral shaped like a lower-order term into that term’s free coefficient, so nothing is counted twice in the same scheme as the power spectrum and shares its infrared resummation. The redshift-space galaxy kernels are derived recursively through fifth order, and through fourth order they reproduce the previously known kernels in exact rational arithmetic. Each two-loop diagram is one of the four topologies drawn below with one pair of bias or velocity coefficients at its vertices; counting the admissible pairs gives 149 diagrams, 46 of which exist only in redshift space. Each diagram splits further into components by powers of the growth ratethe rate at which density fluctuations grow as the universe expands, which sets the size of the galaxy velocities that distort the map $f$ and of $\mu$, the cosine of the angle between $\mathbf{k}$ and the line of sight; in all, 474 components of 136 diagrams are tabulated, each with its numerical error.

The four topologies of the two-loop galaxy power spectrum. Each violet dot is a redshift-space galaxy kernel $Z_n^{\rm gal}$, its subscript giving the perturbative order, and each internal line is one linear power spectrum; the short external lines carry the observed wavenumber $\mathbf{k}$. A closed loop that begins and ends on the same dot is a pair of intermediate modes $(\mathbf{q},-\mathbf{q})$ contracted inside one kernel, a “tadpole”; $P_{33,\mathrm{I}}$ is the only topology without one. In $P_{15}$ and $P_{33,\mathrm{II}}$ the line joining the two dots is not integrated over and gives an overall factor $P_{\rm lin}(k)$. Filling the dots with pairs of bias and velocity coefficients gives 35, 42, 36 and 36 diagrams for the four topologies, 149 in all. (Redrawn from Figure 1 of Paper I.)
The hard part is evaluating the diagrams, because the terms cancel strongly: long-wavelength (soft) contributions cancel between diagrams, short-wavelength (ultraviolet) pieces cancel against counterterms, and inside one diagram the kernel is a sum of individually huge terms of alternating sign. For one matter kernel expanded in monomials the measured cancellation is a factor of $1.3\times10^{17}$, more digits than double-precision arithmetic holds. Each diagram’s summed kernel is therefore integrated directly, and every tabulated component is evaluated twice, in double and in extended precision. The two evaluations must agree to one part in a million; a component that fails this test is re-evaluated, or else its measured round-off is added to its stated error. The pure matter integrands, once the linear spectrum is expanded in power laws, reduce to a two-loop master integralone of a small set of standard integrals to which every integral in a family reduces through algebraic identities of particle physics called the kite, known in closed form for almost every combination of powers on its five lines. One combination that the power spectrum needs, unit powers on two non-adjacent lines, was known only through series; it is obtained here in closed form in three dimensions, complete except for two single sums, with the derivation and its numerical checks in the supplement (Section S3).
The largest of the bare tables (the tabulated components before any subtraction) are those in which both loops close on themselves at composite bias operators. Most of each such table is a contact term, a piece with the shape of a lower-order term, and the renormalization conditions on the lower-order bias parameters remove it. The counterterm operators of the power spectrum absorb the remaining ultraviolet sensitivity, whose leading shape is $k^2 P_{\rm lin}(k)$. The finite part of each table is then fixed by one of two linear subtractions; in the one illustrated below, the pivot condition, the subtracted table must vanish at $k_* = 0.2\,h\,{\rm Mpc}^{-1}$, which moves a $k^2 P_{\rm lin}$-shaped piece out of the table and into the counterterm coefficient.

The pivot subtraction on one two-loop table. The table is the density component of one $P_{15}$ diagram, the linear bias $b_1$ paired with the fifth-order coefficient $\gamma_{41}$, at redshift 0.61, plotted against wavenumber. Orange circles: the bare table as integrated. Gray squares: the part of it with the counterterm’s shape, $\alpha\,k^2 P_{\rm ref}(k)$ with $\alpha = 99.0\,({\rm Mpc}/h)^2$ and $P_{\rm ref}$ a fixed reference linear spectrum. Violet triangles: what is left after that part is subtracted, fixed to vanish at the pivot $k_* = 0.2\,h\,{\rm Mpc}^{-1}$ (stars). The remainder is an order of magnitude smaller than the bare table; its small wiggle near $k_*$ is the acoustic oscillation of $P_{\rm ref}$, which the smooth two-loop table does not share. Most of this bare table thus has the counterterm’s shape and is absorbed into a fitted coefficient, and only the subtracted remainder is a prediction. The numerical errors of the points are smaller than the symbols. (Re-plotted from Figure 3 of Paper I.)
The infrared resummation is extended to two loops, with the damping of the acoustic feature matched to the loop order: the linear, one-loop and two-loop oscillatory parts are damped by $e^{-x}(1+x+x^2/2)$, $e^{-x}(1+x)$ and $e^{-x}$, with $x = k^2\Sigma^2_{\rm tot}$. Here $\Sigma_{\rm tot}$ is the typical relative displacement of two points one acoustic scale apart, computed from the smooth linear spectrum rather than fitted, and the one-loop damping acts inside the loop integrals.
The same methods give the one-loop bispectrum. It has four loop topologies and 568 coefficient functions, the loop integrals that multiply each product of bias and velocity coefficients, integrated on a grid of 786 triangles. The bispectrum model has 17 counterterm coefficients and 19 stochastic parameters, three of the latter shared with the power spectrum. On equilateral triangles at small $k$, at the fiducial bias values, the topology $B_{321,\mathrm{II}}$, which carries a tadpole, is a hundred times larger than $B_{321,\mathrm{I}}$, which does not. On those triangles about 99 percent of the larger topology is a contact term, which renormalization absorbs into the lower-order bias coefficients. The one-loop trispectrum is organized into nine contraction types with up to 5209 coefficient functions and five further stochastic parameters. The contact parts of its self-looped kernels are subtracted, its counterterms are left to future work, and all nine types are evaluated on the 2400 parallelogram configurations that enter the covariancethe table of expected errors on the measured statistics and of the correlations between those errors, needed to weigh data against model of the power spectrum.
Several parts of the calculation remain open. Among them, the counterterms of the one-loop trispectrum remain to be computed; the trispectrum’s contribution to the power-spectrum covariance has been summed for only two of the nine types; the trispectrum is not yet infrared-resummed, an omission bounded at a few percent of its leading term near the acoustic feature; and twelve two-loop components are so far renormalized only by the pivot condition. Comparisons with simulations go through the KITE code, described next. Going to two loops does not extend the reach of the expansion, which still ends near $k \approx 0.2\,h\,{\rm Mpc}^{-1}$, where random galaxy motions along the line of sight become nonperturbative at any loop order. The two-loop term instead reduces the theoretical error inside that range; over $k \simeq 0.1$ to $0.2\,h\,{\rm Mpc}^{-1}$ the loop terms amount to roughly a tenth to a fifth of $P$ at the fiducial point.
Paper II: the KITE code
Full-shape analyses of BOSS and DESI data run on public one-loop codes such as CLASS-PT, PyBird, velocileptors and FOLPSν. These typically evaluate the redshift-space power spectrum at one loop for each cosmology the sampler visits, fit its multipoles (sometimes with a few bispectrum bins at tree level), use the size of the two-loop term only to decide where to stop in $k$, and integrate out the bias and counterterm coefficients analytically under Gaussian priors. KITE, the Kernel-Integral Tensor Engine, is a full-shape likelihood built to put the calculation of Paper I to work on DESI-class data. Precomputing the loop integrals is what makes that affordable, but it is the enabling step rather than the point. The upgrades, and why each helps, are these.
One order higher, three statistics, one calculation. KITE fits the power spectrum, bispectrum and trispectrum together in a single likelihood, with the power spectrum at two loops and the bispectrum and trispectrum at one loop. The extra order pushes the theory error below the statistical error of current surveys over the range of scales they use, and the higher statistics carry information about bias and primordial non-Gaussianity that the power spectrum alone cannot separate. Because all three come from one operator basis, one renormalization scheme and one infrared resummation, each bias or counterterm coefficient means the same thing in every statistic, the bispectrum’s counterterms are tied to the power spectrum’s, and the acoustic feature is damped at the order each term requires. Fitting them jointly then constrains the shared nuisance parametersparameters the model needs but nobody sets out to measure, such as bias, counterterm and shot-noise coefficients; they are fitted and then integrated out from several directions at once instead of leaving each statistic to absorb them separately.
Tables with a stated precision. No loop integral is evaluated during a fit. Every perturbative ingredient is a cosmology-independent table, computed once: the kernel integrals against a fixed set of basis functions for the linear spectrum. A sampled cosmology enters through its coefficients on that basis and the prediction is a contraction of the tables; renormalization, infrared resummation and the Alcock–Paczynski remappingthe rescaling of distances when the data are converted to comoving coordinates with a cosmology other than the true one are linear operations on them. This is what makes a two-loop model fast enough to sit inside a sampler. It also makes the numerics auditable: each two-loop entry is computed in two floating-point precisions and accepted only if they agree to a fixed tolerance, entries that fail are recomputed at higher precision or carried with their error, and those errors are translated into bounds on how far they could move each cosmological parameter, against a requirement fixed before any fit. A per-cosmology integration gives no such statement.
Nuisance parameters integrated out without projection effects. The cosmological parameters and a few nonlinear bias and shot-noise parameters per redshift bin are sampled; the several hundred coefficients that enter the model linearly (bias combinations, counterterms, stochastic amplitudes of all the statistics) are integrated out in one closed-form step. In one-loop analyses this marginalization leaves behind a prior-volume termthe factor left when nuisance parameters are integrated out analytically; it varies with cosmology and can pull amplitude-like parameters away from the full posterior’s peak, the “projection effect” of full-shape analyses that varies with cosmology and is a known source of shifts in parameters such as the fluctuation amplitude. KITE chooses the form of the marginal so that the answer does not depend on how the linear parameters are defined, which removes that term, and reports the conventional form beside it so the difference is visible.
Shot noise and windows treated as the data see them. A survey’s footprint and selection leave an imprint, the window, on every measured statistic. KITE passes the stochastic (shot-noise) terms of the bispectrum through that window exactly as the data are, instead of adding them as a constant, so the model and the measurement are compared like with like down to the noise; the un-windowed version is kept as a control. The bispectrum is measured with a window-free estimator and modeled by the exact average over each bin, with no mixing matrix, and its covariance is built by Monte Carlo through the same window and estimator with a separate allowance for theory error.
Exact posterior, blind fits, choices fixed in advance. The sampled parameters are explored by delayed-acceptance Hamiltonian Monte Carlo: trajectories and a first acceptance test use a fast differentiable approximation to the linear spectrum, and a second test calls the Boltzmann code, so the chains target the exact posterior whatever the quality of the approximation, which affects only the acceptance rate. Runs resume bitwise and reproduce across hardware. Survey fits are blinded in code, by concealed offsets on the cosmological parameters drawn from a seed held outside the fitting team, and the free parameters, scale cuts, priors and acceptance thresholds are fixed by rules exercised on public simulations before any survey posterior is read, so that no analysis choice can be tuned, even unconsciously, toward an expected answer.
The papers
- Toward precision galaxy clustering in redshift space: the next loop order (PDF) — Matthew D. Schwartz, Mikhail M. Ivanov and Siddharth Mishra-Sharma; Paper I. A preliminary version of the paper.
- KITE: a two-loop galaxy clustering pipeline at stated precision — the same three authors; Paper II, in preparation.
- Supplementary Information for Paper II (with the paper, not yet posted) — the full list of tables, the longer derivations and the run-control details.
References
| M. H. Goroff, B. Grinstein, S.-J. Rey and M. B. Wise, Coupling of modes of cosmological mass density fluctuations, Astrophys. J. 311 (1986) 6 | the recursion for the perturbative kernels $F_n$ of the matter density |
| N. Kaiser, Clustering in real space and in redshift space, Mon. Not. Roy. Astron. Soc. 227 (1987) 1 | the leading redshift-space distortion from galaxy velocities |
| P. McDonald and A. Roy, Clustering of dark matter tracers: generalizing bias for the coming era of precision LSS, JCAP 08 (2009) 020 | the general bias expansion relating the galaxy density to the matter density |
| D. Baumann, A. Nicolis, L. Senatore and M. Zaldarriaga, Cosmological non-linearities as an effective fluid, JCAP 07 (2012) 051 | recast the loop expansion as an effective field theory with counterterms |
| J. J. M. Carrasco, M. P. Hertzberg and L. Senatore, The effective field theory of cosmological large scale structures, JHEP 09 (2012) 082 | the effective field theory of large-scale structure, with counterterm coefficients fitted to data |
| J. J. M. Carrasco, S. Foreman, D. Green and L. Senatore, The Effective Field Theory of Large Scale Structures at two loops, JCAP 07 (2014) 057 | the two-loop matter power spectrum in the effective theory, in real space |
| L. Senatore and M. Zaldarriaga, The IR-resummed Effective Field Theory of Large Scale Structures, JCAP 02 (2015) 013 | resummation to all orders of the bulk flows that smear the acoustic feature |
| M. M. Ivanov and S. Sibiryakov, Infrared Resummation for Biased Tracers in Redshift Space, JCAP 07 (2018) 053 | the same resummation for galaxies in redshift space |
| A. Chudaykin, M. M. Ivanov, O. H. E. Philcox and M. Simonović, Nonlinear perturbation theory extension of the Boltzmann code CLASS, Phys. Rev. D 102 (2020) 063533 | CLASS-PT, one of the public one-loop codes used in the BOSS and DESI fits |
| G. D’Amico, L. Senatore and P. Zhang, Limits on wCDM from the EFTofLSS with the PyBird code, JCAP 01 (2021) 006 | PyBird, another of the public one-loop codes named beside CLASS-PT |
| S.-F. Chen, Z. Vlah and M. White, Consistent Modeling of Velocity Statistics and Redshift-Space Distortions in One-Loop Perturbation Theory, JCAP 07 (2020) 062 | velocileptors, the third public one-loop code named in Paper II |
| T. Nishimichi, G. D’Amico, M. M. Ivanov, L. Senatore, M. Simonović, M. Takada, M. Zaldarriaga and P. Zhang, Blinded challenge for precision cosmology with large-scale structure: results from effective field theory for the redshift-space galaxy power spectrum, Phys. Rev. D 102 (2020) 123541 | the PT challenge simulations that Paper II fits as an external test |
| N. A. Maksimova, L. H. Garrison, D. J. Eisenstein, B. Hadzhiyska, S. Bose and T. P. Satterthwaite, AbacusSummit: a massive set of high-accuracy, high-resolution N-body simulations, Mon. Not. Roy. Astron. Soc. 508 (2021) 4017 | the public simulation suite behind the light-cone mock of Paper II’s joint fit |
| P. Taule and M. Garny, The two-loop power spectrum in redshift space, JCAP 11 (2023) 078 | the two-loop matter spectrum in redshift space and its $k\approx0.2\,h\,{\rm Mpc}^{-1}$ limit |
| M. M. Ivanov, Galaxy Power Spectrum at Two-Loop Order: Implications for Weak Lensing Surveys and New Physics, arXiv:2606.30713 (2026) | the two-loop power spectrum of biased tracers in real space |
| T. Bakx, M. Garny, H. Rubira and Z. Vlah, Galaxy bias renormalization: Two-loop Power Spectrum, One-loop Trispectrum and Bispectrum, arXiv:2606.31280 (2026) | fifth-order bias renormalization of all three statistics, in real space |