Higher-Order Fourier Analysis

• Higher-order Fourier analysis is a framework that extends classical Fourier methods to reveal complex correlations using Gowers norms and algebraic objects like nilsequences.
• It integrates concepts from additive combinatorics, ergodic theory, and number theory, applying inverse theorems and regularity lemmas to decompose and understand function behavior.
• Recent developments extend these methods to quantum information and algorithmic testing, showcasing the framework’s broad interdisciplinary applications.

Higher-order Fourier analysis is the study of structured-versus-random dichotomies for functions on finite abelian groups and compact abelian groups beyond the classical first-order dichotomy captured by the usual Fourier transform. Initiated by Gowers and developed through inverse theorems, regularity lemmas, nilspace theory, and polynomial phase methods, it uses the Gowers norms $U^k$ to detect higher-order correlations and explains why a large $U^{k+1}$-norm is governed by algebraic objects such as polynomial phases of degree $k$, nilsequences, or nilspace polynomials (Szegedy, 2012, Candela et al., 21 Jan 2025). The subject grew out of Gowers’s Fourier-analytic proof of Szemerédi’s theorem and now interfaces with additive combinatorics, ergodic theory, number theory, theoretical computer science, and quantum information (Labib, 2021).

1. Classical Fourier analysis and the higher-order shift

Classical Fourier analysis decomposes a bounded function on an abelian group into characters, and in this regime the $U^2$-norm is already Fourier-theoretic: for finite abelian groups one has $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$ (Candela et al., 21 Jan 2025). Higher-order Fourier analysis begins where this description ceases to be adequate, namely when one seeks to control configurations and correlations invisible to linear characters but still encoded by iterated multiplicative differences and cube averages (Szegedy, 2012).

The central qualitative principle is that higher-order structure is not exhausted by periodicity or linear bias. In the formulations developed over finite fields, compact abelian groups, and nilspaces, the relevant structured objects include additive or non-classical polynomial phases, nilsequences, and nilspace polynomials (Bhattacharyya et al., 2015, Szegedy, 2012). In the quadratic case this shift is already substantial: the inverse problem for $U^3$ leads to quadratic phases and, in one algebraic formulation, to quadratic nil-morphisms into 2-step nil-manifolds (Szegedy, 2010).

This change of viewpoint was driven by additive-combinatorial problems. The framework is described as lying at the heart of modern additive combinatorics and underpinning Szemerédi’s theorem, while later work extended it to list decoding, algorithmic decomposition, testing, multiplicative functions, spherical configuration problems, and quantum-information questions (Candela et al., 21 Jan 2025, Bhattacharyya et al., 2015).

2. Uniformity norms, inverse theorems, and regularity

For a bounded function $f:\mathbb K^n\to\mathbb C$ on a finite field $\mathbb K$, the Gowers norm is defined by

$$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$

where $C$ denotes complex conjugation and $U^{k+1}$0 (Bhattacharyya et al., 2015). On compact abelian groups the same norm is expressed via iterated differences,

$U^{k+1}$1

with $U^{k+1}$2 (Szegedy, 2012).

The inverse problem asks what large $U^{k+1}$3 means structurally. Over finite fields, Tao–Ziegler’s inverse theorem states that if $U^{k+1}$4 and $U^{k+1}$5, then $U^{k+1}$6 correlates with an additive polynomial $U^{k+1}$7 of degree at most $U^{k+1}$8, in the sense that $U^{k+1}$9 (Bhattacharyya et al., 2015). In the prime-field setting, when $k$0 the correlating object may be taken to be a classical polynomial of degree at most $k$1, but in low characteristic $k$2 one must allow non-classical phase-polynomials taking values in $k$3 (Candela et al., 2021). A recurrent technical misconception is therefore that classical polynomial phases always suffice; the low-characteristic theory explicitly requires non-classical phases (Candela et al., 2021).

Regularity theory provides the complementary decomposition statement. In one finite-field formulation, every bounded $k$4 can be decomposed as

$k$5

where $k$6 is measurable with respect to a polynomial factor of bounded complexity and degree, $k$7 has small $k$8-norm, and $k$9 has small $U^2$0-norm (Bhattacharyya et al., 2015). In the compact-group and ultraproduct setting, one obtains a nilspace-polynomial regularity lemma and an exact decomposition in which the uniform component has vanishing $U^2$1-norm on the limiting object (Szegedy, 2012). The subject’s basic analytic architecture is thus the pairing of an inverse theorem with a regularity or structure theorem.

3. Structured objects: additive polynomials, nilspaces, and nilsequences

One finite-field language for higher-order structure is the language of additive polynomials. A function $U^2$2 has additive degree at most $U^2$3 if

$U^2$4

where $U^2$5 (Bhattacharyya et al., 2015). Classical polynomial phases are special cases: $U^2$6 for a $U^2$7-polynomial $U^2$8 of total degree at most $U^2$9 (Bhattacharyya et al., 2015). In low characteristic these sit inside the broader class of additive, or non-classical, polynomials (Bhattacharyya et al., 2015).

A more geometric and global language is that of nilspaces. A compact $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$0-step nilspace is a compact space endowed with cube sets $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$1 satisfying composition, ergodicity, and $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$2-uniqueness, and every compact $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$3-step nilspace is an inverse limit of finite-dimensional ones built as iterated abelian bundles over lower-step factors (Szegedy, 2012). In this formulation, structured functions are nilspace-polynomials: compositions $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$4 where $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$5 is a continuous morphism into a compact $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$6-step nilspace and $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$7 is bounded and Lipschitz (Szegedy, 2012). This gives an algebraic interpretation of higher-order Fourier analysis in terms of continuous morphisms between compact $$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$8-step nilspaces (Szegedy, 2012).

In the characteristic-$$\|f\|_{U^2}^4=\sum_{\chi\in\widehat G}|\widehat f(\chi)|^4$$9 setting, Candela, González-Sánchez, and Szegedy introduced $U^3$0-homogeneous nilspaces as the nilspace class adapted to vector spaces over $U^3$1 (Candela et al., 2021). For a filtered group $U^3$2, the associated group nilspace is $U^3$3-homogeneous if and only if for all $U^3$4 and all $U^3$5, one has $U^3$6 (Candela et al., 2021). Every finite $U^3$7-homogeneous nilspace is the image, under a nilspace fibration, of a member of a simple family of filtered finite abelian $U^3$8-groups, and this structure theorem yields a new proof of the Tao–Ziegler inverse theorem for Gowers norms on $U^3$9 (Candela et al., 2021).

At the quadratic level, Szegedy’s algebraic theory describes structure through quadratic nil-morphisms into 2-step nil-manifolds. In that setting, a bounded function on a finite abelian group decomposes as

$f:\mathbb K^n\to\mathbb C$0

where $f:\mathbb K^n\to\mathbb C$1 is a quadratic nil-morphism, $f:\mathbb K^n\to\mathbb C$2 is a bounded-complexity continuous function on a 2-step nil-manifold, $f:\mathbb K^n\to\mathbb C$3 is small, and $f:\mathbb K^n\to\mathbb C$4 is small in $f:\mathbb K^n\to\mathbb C$5 (Szegedy, 2010). This provides a quadratic inverse theorem for $f:\mathbb K^n\to\mathbb C$6 in which correlation is with a nil-manifold model rather than merely with an explicit phase function (Szegedy, 2010).

4. Extensions of the ambient domain

Although much of the early development was over prime fields, the framework extends to arbitrary finite fields. For $f:\mathbb K^n\to\mathbb C$7 one fixes an $f:\mathbb K^n\to\mathbb C$8-basis and uses the trace map $f:\mathbb K^n\to\mathbb C$9 to pass between $\mathbb K$0-valued functions and torus-valued phases (Bhattacharyya et al., 2015). Bhattacharyya and Bhowmick showed that the inverse theorem, polynomial regularity lemma, counting and equidistribution machinery, and applications to affine-invariant properties can be transported to general finite fields, with new ingredients including equidistribution on $\mathbb K$1-affine subspaces and preservation of locally characterized properties (Bhattacharyya et al., 2015). A key obstruction in nonprime fields is that the naive derivative argument fails: for example, $\mathbb K$2 has trivial second derivatives but degree $\mathbb K$3 (Bhattacharyya et al., 2015).

The theory also has a compact-group and ultraproduct form. In this setting one defines higher-order $\mathbb K$4-algebras $\mathbb K$5 on an ultraproduct group, proves that $\mathbb K$6 is a norm on $\mathbb K$7, and shows that separable sub-$\mathbb K$8-algebras of $\mathbb K$9 are generated by continuous nilspace factors (Szegedy, 2012). This exact limit theory supplies inverse theorems, regularity lemmas, and a limit theory for bounded functions on abelian groups analogous in spirit to graph limit theory (Szegedy, 2012).

A geometric extension replaces cubes in the ambient vector space by cubes constrained to lie on a quadratic variety. In the spherical finite-field setting, for $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$0 one defines local Gowers norms by averaging only over Gowers cubes contained in $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$1 (Sun, 2023). Sun proved a quantitative equidistribution theorem for polynomial sequences on nilmanifolds averaged along spheres: if $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$2 is not $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$3-equidistributed on the lifted sphere, then there exists a nontrivial horizontal character $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$4 such that $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$5 is constant modulo $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$6 on that sphere (Sun, 2023). The corresponding spherical inverse theorem asserts that if the $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$7-th spherical Gowers norm is at least $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$8 and $$U^d(f)^{2^d} = \mathbb E_{x,h_1,\dots,h_d\in\mathbb K^n} \prod_{\varepsilon\in\{0,1\}^d} C^{|\varepsilon|} f(x+\varepsilon\cdot h),$$9, then the function correlates on the sphere with a $C$0-periodic $C$1-step nilsequence of bounded complexity (Sun, 2023).

Higher-order notions also extend beyond functions to singular measures on $C$2 or $C$3. Carnovale introduced a measure $C$4 and a uniformity norm $C$5 for measures, together with the $C$6th-order Fourier dimension, defined through decay of the Fourier transform of the higher-order difference measure (Carnovale, 2013). The main result is that sufficiently strong $C$7th-order Fourier decay controls the rate at which $C$8, where $C$9 is an approximation to $U^{k+1}$00 (Carnovale, 2013). This shows that higher-order Fourier analysis can register interactions among frequency components not visible from ordinary Fourier decay or $U^{k+1}$01 information alone (Carnovale, 2013).

5. Applications across combinatorics, computer science, number theory, and quantum information

In theoretical computer science, the general-field extension of higher-order Fourier analysis yields three applications stated in uniform finite-field form. For any fixed finite field $U^{k+1}$02, the list decoding radius of the generalized Reed–Muller code equals the minimum distance of the code; for any fixed finite field $U^{k+1}$03, there is a deterministic polynomial-time algorithm, running in time $U^{k+1}$04 up to factors depending only on $U^{k+1}$05, that decides whether a degree-$U^{k+1}$06 polynomial is $U^{k+1}$07-structured and finds such a decomposition if it exists; and every $U^{k+1}$08-lightly locally characterized affine-invariant property admits a proximity-oblivious one-sided tester making $U^{k+1}$09 queries (Bhattacharyya et al., 2015). These results are all derived from the same package of inverse theorems, regularity, equidistribution, and degree-preservation (Bhattacharyya et al., 2015).

In number theory, Frantzikinakis and Host proved a structure theorem for bounded multiplicative functions: every such function decomposes into an approximately periodic term and another with small Gowers uniformity norm of arbitrary degree, with an additional small $U^{k+1}$10 error after refinement (Frantzikinakis et al., 2014). They used this to characterize when the Gowers norms of a bounded multiplicative function are zero, to prove asymptotic orthogonality to irrational nilsequences, to establish Chowla-type zero-mean results for certain quadratic forms in two variables, and to obtain partition-regularity results for homogeneous quadratic equations in three variables, including $U^{k+1}$11 and $U^{k+1}$12 (Frantzikinakis et al., 2014).

In quantum information, stabilizer states over prime-dimensional qudits can be written exactly as quadratic phase functions on affine subspaces. Concretely, if $U^{k+1}$13 is an $U^{k+1}$14-qudit stabilizer state over $U^{k+1}$15, then there exist an affine subspace $U^{k+1}$16 and a quadratic polynomial $U^{k+1}$17 such that

$U^{k+1}$18

and in the qubit case one similarly obtains a $U^{k+1}$19-valued quadratic form on an affine subspace (Labib, 2021). This identifies stabilizer states with nonclassical quadratic phase functions on affine subspaces and permits explicit use of higher-order Fourier-analytic tools. One consequence is that for any prime $U^{k+1}$20, the $U^{k+1}$21-qudit magic state $U^{k+1}$22 has stabilizer rank $U^{k+1}$23 (Labib, 2021).

The spherical program applies higher-order Fourier analysis to geometric Ramsey theory over finite fields. For a spherical configuration $U^{k+1}$24 of complexity at most $U^{k+1}$25, with $U^{k+1}$26 and $U^{k+1}$27 sufficiently large with respect to $U^{k+1}$28 and $U^{k+1}$29, Sun proved that every set $U^{k+1}$30 with $U^{k+1}$31 contains at least $U^{k+1}$32 congruent copies of $U^{k+1}$33 (Sun, 2023). The novelty stated for this argument is that it avoids the use of harmonic analysis and replaces it by the theory of spherical higher-order Fourier analysis developed in the preceding parts of the series (Sun, 2023).

Algorithmic and spectral developments have also appeared inside the subject itself. A recent spectral framework turns quadratic higher-order Fourier analysis into an eigenvalue problem by applying invariant operators to the matrix $U^{k+1}$34 and studying the spectrum of the resulting Hermitian operator (Candela et al., 21 Jan 2025). In the quadratic case this yields a spectral inverse theorem, a spectral regularity theorem, and an explicit $U^{k+1}$35-regularization algorithm with total time $U^{k+1}$36 (Candela et al., 21 Jan 2025).

6. Quantum higher-order Fourier analysis and current directions

A noncommutative extension, termed quantum higher-order Fourier analysis, replaces functions on a finite abelian group by linear operators on an $U^{k+1}$37-qudit Hilbert space, ordinary translations by conjugation with Weyl operators, and Gowers norms by quantum uniformity measures $U^{k+1}$38 (Bu et al., 21 Aug 2025). If $U^{k+1}$39 is diagonal in the computational basis with diagonal entries $U^{k+1}$40, then $U^{k+1}$41, so the classical theory is recovered on diagonal operators (Bu et al., 21 Aug 2025). The principal analytic theorem is a characterization of the Clifford hierarchy:

$U^{k+1}$42

giving a necessary and sufficient analytic condition for membership in the $U^{k+1}$43th level (Bu et al., 21 Aug 2025). The same framework yields overlap bounds with lower hierarchy levels and a property-testing interpretation based on sampling phase-space derivatives (Bu et al., 21 Aug 2025).

Several present directions follow directly from the current literature. The stabilizer-rank work suggests studying more refined measures of non-stabilizerness via higher Gowers $U^{k+1}$44 norms and inverse theorems for cubic or quartic phases, and it raises the possibility of super-linear or even exponential lower bounds on decompositions into full-space stabilizer states through stronger inverse theorems for $U^{k+1}$45 with $U^{k+1}$46 (Labib, 2021). The quantum higher-order Fourier-analysis framework, in turn, formulates a quantum analogue of the classical inverse theorem: if $U^{k+1}$47, then one conjectures that $U^{k+1}$48 overlaps with some element of $U^{k+1}$49 by an amount bounded below by $U^{k+1}$50 (Bu et al., 21 Aug 2025).

The cumulative picture is that higher-order Fourier analysis is not a single formalism but a family of tightly related ones: additive-polynomial over finite fields, nilspace and nilmanifold on compact or ergodic systems, local and spherical on quadratic varieties, spectral in algorithmic quadratic analysis, and noncommutative in the quantum setting (Szegedy, 2012, Candela et al., 2021, Sun, 2023, Bu et al., 21 Aug 2025). A plausible implication is that the subject’s central invariant is less the ambient domain than the existence of an appropriate notion of higher-order derivative, cube, and structured phase against which uniformity can be tested.