---
title: NIP Arithmetic Regularity Lemma
url: https://www.emergentmind.com/topics/nip-arithmetic-regularity-lemma
type: topic
---

# NIP Arithmetic Regularity Lemma

Searching arXiv for recent and foundational papers on NIP arithmetic regularity, stable arithmetic regularity, and classical arithmetic regularity lower bounds.
In the literature summarized here, the phrase **NIP arithmetic regularity lemma** is best understood as naming a family of tame arithmetic regularity results for subsets of groups whose translate families have bounded VC-dimension, together with closely related definable regularity theorems in NIP theories. Their common objective is to replace the unrestricted Fourier-analytic regularity paradigm by bounded-complexity structured approximations—typically via subgroups, Bohr neighborhoods, or definable partitions—while forcing almost all relevant cells to be nearly empty or nearly full. The subject is usually read against Green’s classical arithmetic regularity lemma, whose unrestricted form is intrinsically tower-complex, whereas bounded VC-dimension, stability, and distality yield much stronger structure and markedly better quantitative behavior [1405.4409, 2509.04271, 1710.02021].

## 1. Classical arithmetic regularity and the motivation for tame variants

Green’s arithmetic regularity lemma, in the finite-field model \(G=\mathbb{Z}_2^n\), concerns bounded functions \(f:G\to[0,1]\) and asks for a subspace \(H\leq G\) such that on most cosets \(H+g\), the restriction \(f|_{H+g}\) has no large nontrivial Fourier coefficient. In the formulation used in the lower-bound literature, a coset \(H+g\) is \(\epsilon\)-regular when
\[
\max_{\eta\notin H^\perp} |\widehat{f|_{H+g}}(\eta)|\le \epsilon,
\]
and \(H\) is \(\epsilon\)-regular for \(f\) if this holds on at least a \((1-\epsilon)\)-fraction of cosets. Green’s upper bound, in the version cited in the finite-field lower-bound paper, is
\[
M(\epsilon)\le \operatorname{twr}(\lceil 1/\epsilon^3\rceil),
\]
where \(\operatorname{twr}(h)\) is a tower of twos of height \(h\) [1405.4409].

The decisive point for later NIP regularity is quantitative: unrestricted arithmetic regularity is genuinely expensive. The lower bound of Hosseini, Lovett, Moshkovitz, and Shapira shows
\[
M(\epsilon)\ge \operatorname{twr}\!\big(\lfloor 1/(16\epsilon)\rfloor\big),
\]
improving Green’s earlier \(\Omega(\log(1/\epsilon))\) tower-height lower bound to \(\Omega(1/\epsilon)\). In other words, even in \(G=\mathbb{Z}_2^n\), tower-type dependence is essentially unavoidable in the unrestricted Fourier-analytic setting [1405.4409].

Subsequent lower-bound work sharpened this contrast. For strong arithmetic regularity partitions, wowzer-type growth is forced; for higher-order arithmetic regularity, tower-type growth remains necessary at least for the linear layer of a quadratic partition [2510.15532]. This quantitative barrier is the main background against which NIP and stable arithmetic regularity are interpreted: the tame hypotheses are not cosmetic, but are the mechanism by which one escapes the worst-case tower and wowzer phenomena.

## 2. The finite-group NIP arithmetic regularity lemma

In the finite-group setting, “NIP” is treated combinatorially via VC-dimension. A subset \(A\subseteq G\) is called \(d\)-NIP when the collection of left translates of \(A\) has VC-dimension less than \(d\). The new proof of the finite-group NIP arithmetic regularity lemma works with the local parameter
\[
d=\max\{VC^\ell_A(A),VC^r_A(A)\},
\]
where \(VC^\ell_B(A)\) and \(VC^r_B(A)\) measure the VC-dimension of the families \(\{xA:x\in B\}\) and \(\{Ax:x\in B\}\), respectively [2509.04271].

The theorem proved there applies to **arbitrary finite groups**. If \(G\) is finite, \(A\subseteq G\) is nonempty, \(\alpha=|A|/|G|\), and \(\epsilon\in(0,1)\), then there exists a Bohr neighborhood
\[
B\subseteq St^r_\epsilon(A)
\]
of complexity \(O_{d,\alpha,\epsilon}(1)\) such that two conclusions hold. First, there is a set \(F\subseteq A\) with \(|F|\le O_{d,\alpha,\epsilon}(1)\) for which
\[
|A\triangle FB|<\epsilon|A|.
\]
Second, if
\[
Z^\ell_\epsilon(A,B)=\{g\in G:\min(|gB\cap A|,\ |gB\setminus A|)\ge \epsilon |B|\},
\]
then
\[
|Z^\ell_\epsilon(A,B)|\le \epsilon |A|.
\]
Thus \(A\) is approximated, up to relative error \(\epsilon\), by a bounded union of left translates of a bounded-complexity Bohr neighborhood, and all but at most \(\epsilon|A|\) left translates \(gB\) are almost homogeneous for \(A\) [2509.04271].

The theorem is phrased in terms of stabilizers:
\[
Stab^\ell_N(A)=\{x\in G:|xA\triangle A|\le N\},\qquad
Stab^r_N(A)=\{x\in G:|Ax\triangle A|\le N\},
\]
with \(St^\bullet_\epsilon(A)=Stab^\bullet_{\epsilon|A|}(A)\). The inclusion \(B\subseteq St^r_\epsilon(A)\) is conceptually important: the structured object controlling \(A\) lies inside an approximate symmetry set of \(A\), not merely inside an unrelated ambient subgroup or progression [2509.04271].

This theorem recovers the earlier Conant–Pillay–Terry finite-group NIP regularity lemma, but with a different proof architecture. The new argument avoids Borel definability and generic compact domination, replacing them with VC-dimension estimates, stabilizer technology, and additive-combinatorial structure theorems [2509.04271].

## 3. Stable arithmetic regularity as the strongest tame regime

Stability is a strictly stronger hypothesis than NIP, and in arithmetic regularity it produces correspondingly stronger conclusions. For \(A\subseteq \mathbb{F}_p^n\), \(k\)-stability means that \(A\) does not contain the additive half-graph pattern
\[
a_i+b_j\in A \iff i\le j.
\]
Under this hypothesis, Terry and Wolf proved a stable arithmetic regularity lemma that eliminates both of the main pathologies of Green’s theorem: tower-type codimension and exceptional cosets [1710.02021].

Precisely, for all \(\epsilon\in(0,1)\), \(k\ge 2\), and primes \(p\), there exists \(n_0=n_0(k,\epsilon,p)\) such that if \(n\ge n_0\), \(G=\mathbb{F}_p^n\), and \(A\subseteq G\) is \(k\)-stable, then there is a subspace \(H\le G\) of codimension
\[
O_k(\epsilon^{-O_k(1)})
\]
such that for **every** \(g\in G\), either
\[
|(A-g)\cap H|\le \epsilon|H|
\quad\text{or}\quad
|H\setminus (A-g)|\le \epsilon|H|.
\]
Equivalently, every coset \(g+H\) has density in \([0,\epsilon]\cup[1-\epsilon,1]\). There are no exceptional cosets, and the codimension bound is polynomial in the uniformity parameter for fixed \(k\) [1710.02021].

This is much stronger than classical Fourier-uniformity on most cosets. It implies that stable sets are approximable by unions of cosets of a bounded-codimension subspace, and the paper also gives robustness under small perturbation. The result is explicitly presented as an arithmetic analogue of the Malliaris–Shelah stable graph regularity lemma [1710.02021].

The stable side also admits an analytic extension beyond indicator sets. For amenable groups, a \(k\)-stable function \(f:G\to[-1,1]\) is shown to be almost constant on all translates of a bounded-complexity \((\delta,U(n))\)-Bohr neighborhood: there exists \(B\) with
\[
\delta^{-1},n\le O_{k,\zeta,\epsilon}(1)
\]
such that \(f\) is \(\zeta(\delta,n)\)-almost \(\epsilon\)-constant on all translates of \(B\). In this formulation the structured object is no longer a subgroup factor but a unitary Bohr neighborhood, and the result is interpreted as approximate uniform Bohr-continuity for stable functions [2401.14363].

## 4. Definable NIP, distal, and hypergraph formulations

A second major line of development treats regularity in definable, measure-theoretic terms. For definable \(k\)-ary hypergraph relations \(R(x_1,\dots,x_k)\) in NIP structures, the regularity statement is not Fourier-analytic but partition-based: one seeks finite definable partitions of the ambient sorts so that outside a small exceptional family of cells, each product cell has density approximately \(0\) or approximately \(1\). In the general NIP case, if \(E(x_1,\dots,x_k)\) is definable and the \(\mu_i\) are generically stable Keisler measures, there are partitions
\[
M^{|x_i|}=\bigcup_{j<K} A_{i,j}
\]
with
\[
K\le (1/\epsilon)^c
\]
such that the union of exceptional cells has product measure at most \(\epsilon\), and on every nonexceptional cell the \(E\)-density is within \(\epsilon\) of either \(0\) or \(1\) [1607.07701].

This result sits between the stable and distal extremes. In the stable case, the exceptional set disappears: \(\Sigma=\emptyset\), so every cell is regular. In the distal case, the conclusion strengthens in a different direction: outside the exceptional set, cells are exactly homogeneous,
\[
X\cap R=\emptyset \quad\text{or}\quad X\subseteq R,
\]
and the formulas defining the cells depend only on the original relation, not on \(\epsilon\) [1607.07701].

A recent higher-arity extension makes this distal picture explicit for strongly \(n\)-distal NIP theories. For a definable \((n+1)\)-ary relation \(\varphi(x_0,\dots,x_n)\), one obtains partitions in the co-directions \(\prod_{j\neq i}M^{x_j}\) such that the total product measure of the cylinder-intersection cells that are **not** \(\varphi\)-homogeneous is less than \(\epsilon\). Under additional hypotheses such as definable Skolem functions, the same conclusion holds for all generically stable measures with uniformly definable partitions [2605.04714].

This strongly \(n\)-distal theory also has a group-theoretic consequence of direct arithmetic relevance. If \(G\) is a definable group and \(\mu\) is a \(G\)-invariant generically stable measure, then \(\mu\) is smooth. By the compact-domination criterion for smooth invariant measures, compact domination holds for definable fsg groups in strongly \(n\)-distal NIP theories, and hence an arithmetic version of the distal regularity lemma holds for definable groups exactly as in the earlier distal literature [2605.04714].

A common source of confusion is terminological. Definable NIP hypergraph regularity is a genuine regularity theory within the NIP world, but it is not the same theorem as the additive-combinatorial arithmetic regularity lemma of Green; its ambient objects, norms, and partition classes are different, even when the philosophical goal—structured approximation plus a small exceptional set—is closely related [1607.07701].

## 5. Proof mechanisms and structural ideas

The new finite-group proof organizes NIP arithmetic regularity around a single principle: **bounded VC-dimension forces many approximate symmetries**. The basic input is Haussler’s packing lemma. Applied to the translate family of \(A\), it yields covering bounds by stabilizers, for example
\[
cov(B:St^\ell_\epsilon(A))\le \left(\frac{30|BA|}{\epsilon|A|}\right)^d
\]
and the analogous right-handed estimate. Thus, bounded VC-dimension implies that many group elements translate \(A\) only slightly in symmetric difference [2509.04271].

The central technical advance is that one no longer needs an actual subgroup inside the stabilizer. Two lemmas show that **arbitrary subsets** of a stabilizer already suffice. If \(X\subseteq Stab^r_N(A)\), then
\[
|Z^\ell_\epsilon(A,X)|\le \frac{2N}{\epsilon},
\]
so most left translates of \(X\) are nearly monochromatic for \(A\). A second lemma, modeled on work of Sisask, shows that if one constructs a large subset \(A'\subseteq A\) and a right stabilizer set \(S\) with
\[
A'\subseteq D\subseteq A'S,
\]
then \(D\) already approximates \(A\) in symmetric difference. This is the key step that allows Bohr neighborhoods and nilprogressions—rather than only subgroups—to act as regularity objects [2509.04271].

Once stabilizers are known to be large, additive-combinatorial structure theorems enter. In the finite-group dense setting, a noncommutative Bogolyubov theorem places a bounded-complexity Bohr neighborhood inside a bounded product of a large symmetric set. In the bounded-tripling setting, the Alon–Fox–Zhao trick first extracts a set of controlled tripling from the stabilizer, and then the Breuillard–Green–Tao approximate-group theorem supplies a coset nilprogression [2509.04271].

At a more abstract level, the common logic of these arguments is still the regularity logic of test families and refinement. The unified abstract regularity lemma of 2026 isolates this mechanism as follows: one chooses structured partitions, assigns local test-function families \(\mathcal F_P\), defines local irregularity by correlation against \(\mathcal F_P\), and refines whenever the average irregularity exceeds \(\epsilon\). The energy increment
\[
\mathcal E_g(\mathcal Q)-\mathcal E_g(\mathcal P)
\]
controls termination after at most \(\lceil \epsilon^{-2}\rceil\) steps. In Green’s arithmetic regularity lemma, the local tests are nontrivial characters on cosets, the structured partitions are coset partitions of subspaces, and refinement is Fourier refinement [2606.06192]. This suggests that later NIP variants differ less in global architecture than in the choice of admissible tests and structured factors.

## 6. Quantitative landscape, higher-order directions, and conceptual boundaries

The quantitative contrast among the main regimes can be summarized concisely.

| Regime | Structured object | Typical conclusion |
|---|---|---|
| Classical Green regularity | Subspace \(H\le \mathbb F_p^n\) | Fourier-uniform on most cosets; tower bounds |
| Finite-group NIP | Bohr neighborhood \(B\subseteq St_\epsilon(A)\) | \(A\approx FB\); most translates nearly homogeneous |
| Stable finite-field regularity | Bounded-codimension subspace \(H\) | Every coset almost empty or almost full |
| Definable NIP hypergraph regularity | Definable partitions | Nonexceptional cells have density near \(0\) or \(1\) |

The unrestricted lower bounds are now very sharp. For Green’s classical arithmetic regularity lemma, tower-type complexity of height linear in \(1/\epsilon\) is necessary [1405.4409]. For strong arithmetic regularity, wowzer-type growth is necessary, and in higher-order arithmetic regularity the linear layer of a quadratic partition still requires tower-type codimension [2510.15532]. Against that background, the polynomial-size bounds for definable NIP hypergraph regularity and the polynomial-codimension/no-exceptional-coset theorem in the stable finite-field model are mathematically substantive rather than cosmetic improvements [1607.07701, 1710.02021].

Higher-order analogues indicate that ordinary NIP is not the end of the story. Terry and Wolf define higher-arity tameness notions such as \(VC_2\)-dimension and \(FOP_2\), and show that if \(A\subseteq \mathbb F_p^n\) has bounded \(VC_2\)-dimension then \(A\) is approximable by a union of atoms of a bounded-complexity high-rank quadratic factor, up to a small proportion of exceptional atoms. If \(A\) omits \(k\)-\(FOP_2\), the exceptional region can be confined to a small number of **linear** atoms, yielding what the paper calls **linear error** [2111.01739]. This suggests that genuinely quadratic regularity is naturally governed by higher-arity tameness rather than by ordinary binary NIP alone.

Several distinctions are therefore essential. The NIP arithmetic regularity lemma is not a single theorem but a cluster of results across finite groups, finite-field models, and definable settings. Stable arithmetic regularity is a special case inside NIP, with stronger conclusions than are presently known in general. Definable NIP hypergraph regularity is adjacent to, but not identical with, additive-combinatorial arithmetic regularity. And the persistent tower and wowzer lower bounds in the unrestricted setting explain why model-theoretic tameness assumptions are central: they are precisely what allows arithmetic regularity to become structurally informative and quantitatively manageable [2509.04271, 2605.04714].

Source: https://www.emergentmind.com/topics/nip-arithmetic-regularity-lemma