Green–Tao Transference Principle

• The Green–Tao transference principle is a mechanism that upgrades dense combinatorial statements to sparse settings using a pseudorandom majorant.
• It employs dense model theorems and relative counting lemmas to transfer Szemerédi’s theorem to sparse structures like the primes.
• The method extends to multidimensional and nonclassical settings, enabling breakthroughs such as arbitrarily long arithmetic progressions in sparse sets.

The Green–Tao transference principle is the mechanism that upgrades dense combinatorial statements, most prominently Szemerédi’s theorem, to sparse settings governed by a pseudorandom majorant. In the form used in the modern exposition of the Green–Tao theorem, it asserts that if a sparse function $f$ is dominated by a measure $\nu$ that behaves like the constant function $1$ on the multilinear statistics relevant to $k$-term arithmetic progressions, then $f$ admits a bounded dense model $g$ with essentially the same progression counts; Szemerédi’s theorem applies to $g$, and a counting lemma transfers the resulting lower bound back to $f$. Applied to a $W$-tricked majorant for the primes, this yields the theorem that the primes contain arbitrarily long arithmetic progressions (Conlon et al., 2014).

1. Relative Szemerédi formulation

For fixed $k \ge 3$ and $\nu$0, write $\nu$1 and $\nu$2, and let $\nu$3 denote normalized averages. The basic progression-counting operator is

$\nu$4

In the cyclic model one writes

$\nu$5

The dense input is Szemerédi’s theorem in weighted form: for every $\nu$6 and $\nu$7 there exists $\nu$8 such that every $\nu$9 with $1$0 satisfies

$1$1

The transference statement is the relative Szemerédi theorem. If $1$2 satisfies the $1$3-linear forms condition and $1$4 obeys $1$5 with $1$6, then

$1$7

and one may take the same constant $1$8 as in the dense theorem. Translated back to $1$9, this yields $k$0 (Conlon et al., 2014).

The conceptual content is that density is not required absolutely; it is enough to have positive density relative to a background measure $k$1 that is pseudorandom at the level of the configurations being counted. This is the precise sense in which the transference principle “upgrades” Szemerédi’s theorem from dense sets to sparse pseudorandom sets.

2. Pseudorandom majorants and the linear forms condition

The majorant $k$2 is required to look like the constant function $k$3 for all multilinear statistics needed to count $k$4-term arithmetic progressions. In the Conlon–Fox–Zhao formulation, pseudorandomness is expressed by a linear forms condition. A representative quantitative version requires that for any finite-complexity system $k$5 of affine-linear forms on $k$6,

$k$7

In the hypergraph formulation, this is equivalent to the statement that the weighted $k$8-uniform $k$9-partite hypergraph built from $f$0 has asymptotically the expected count of every subhypergraph of the $f$1-blow-up of the simplex $f$2 (Conlon et al., 2014).

In the original Green–Tao proof, pseudorandomness was encoded more heavily: besides linear forms estimates, one imposed correlation conditions such as

$f$3

for bounded $f$4 and small distinct shifts $f$5, together with small correlations against dual functions associated to $f$6. The CFZ approach shows that, for the relative Szemerédi theorem, the linear forms condition alone is sufficient; the separate correlation condition can be omitted (Conlon et al., 2014).

A common misconception is to interpret pseudorandomness as pointwise resemblance to a random set. The transference principle uses a narrower and more structural notion: $f$7 must behave like $f$8 only on the multilinear averages that arise in the counting problem. The data make this explicit: these conditions ensure that $f$9 behaves like $g$0 at the level of all multilinear statistics needed to count $g$1-APs.

3. Dense models, cut norms, and relative counting

The CFZ implementation is organized around a dense model theorem and a counting lemma. For an $g$2-linear form one defines a cut norm; in the arithmetic model,

$g$3

For the $g$4-AP problem one takes $g$5.

The dense model theorem states that for every $g$6 there exists $g$7 such that if $g$8 satisfies $g$9, then every $g$0 with $g$1 admits a bounded model $g$2 with

$g$3

The relevant test functions are generalized convolutions of indicator functions; this family is closed under multiplication, so the associated dual norm has a multiplicatively closed unit ball (Conlon et al., 2014).

The second ingredient is the relative counting lemma. In its hypergraph form, if $g$4 are weighted $g$5-uniform $g$6-partite hypergraphs with $g$7, $g$8, $g$9 satisfying the $f$0-linear forms condition, and $f$1, then the simplex counts differ by $f$2. A key input is densification, combined with repeated Cauchy–Schwarz and strong linear forms lemmas that allow deletion of a $f$3-factor at $f$4 cost (Conlon et al., 2014).

To recover arithmetic progressions, one builds hypergraphs from the linear forms associated to $f$5-APs. For $f$6, define

$f$7

and set $f$8, $f$9. The relative simplex counting lemma then yields

$W$0

Combining this with dense Szemerédi for $W$1 proves the relative theorem.

4. The primes: $W$2-trick, enveloping sieve, and transference

For the primes, the sparse object is modeled by the von Mangoldt function

$W$3

The prime number theorem is equivalent to $W$4.

The $W$5-trick removes local congruence biases. Let $W$6 slowly and

$W$7

For $W$8, define

$W$9

The normalization $k \ge 3$0 makes the mean close to $k \ge 3$1 among residue classes coprime to $k \ge 3$2 (Conlon et al., 2014).

A pseudorandom majorant is then constructed by an enveloping sieve. With a smooth cutoff $k \ge 3$3 supported on $k \ge 3$4, define

$k \ge 3$5

and

$k \ge 3$6

For suitable parameters, $k \ge 3$7 has mean $k \ge 3$8 and satisfies the $k \ge 3$9-linear forms condition. The crucial linear forms estimate asserts that for fixed linear maps $\nu$00 with no two proportional, averages of products of $\nu$01 over boxes are asymptotic to the expected main term; this supplies the required pseudorandomness of $\nu$02 (Conlon et al., 2014).

One then takes a function below $\nu$03, for example

$\nu$04

with $\nu$05 and $\nu$06. The dense model theorem produces $\nu$07 with the same mean and small cut-norm error. The relative counting lemma gives $\nu$08, and weighted Szemerédi gives $\nu$09. Hence $\nu$10. Since $\nu$11 is supported on $\nu$12, the contribution from $\nu$13 is negligible, so a positive lower bound forces the existence of $\nu$14 with $\nu$15 and

$\nu$16

As $\nu$17 is arbitrary, the primes contain arbitrarily long arithmetic progressions (Conlon et al., 2014).

5. Competing formulations and quantitative refinements

The original Green–Tao approach was organized around Gowers uniformity norms, dual functions, and a relative hypergraph removal lemma. For $\nu$18, the $\nu$19 norm on $\nu$20 is

$\nu$21

Control of $\nu$22, or of correlations against $\nu$23-dual functions, suffices to ensure that progression-counting multilinear forms behave as in the dense model. The CFZ method replaces this by cut norms and relative hypergraph counting; in the graph case, the cut norm is weaker than $\nu$24, and the argument avoids the full Gowers machinery (Conlon et al., 2014).

A second reformulation is Zhao’s arithmetic transference proof, which applies the Green–Tao–Ziegler dense model theorem directly with a discrepancy or cut-type norm adapted to $\nu$25-APs, then combines it with a counting lemma and Szemerédi’s theorem as a black box. In that setup the relative Szemerédi theorem inherits the same constant $\nu$26 as the dense weighted theorem, rather than a weaker constant emerging from hypergraph removal (Zhao, 2013).

Quantitative refinements sharpened the transference pipeline. Rimanić and Wolf combined quantified relative Szemerédi with optimized enveloping sieve weights to obtain density bounds for $\nu$27-AP-free subsets of the primes for $\nu$28 (Rimanic et al., 2017). A later quantitative advance establishes that if a set $\nu$29 of relative density $\nu$30 within the primes up to $\nu$31 contains no nontrivial arithmetic progressions of length $\nu$32, then

$\nu$33

for some $\nu$34. The main new ingredients are a quasipolynomial inverse theorem for unbounded functions and a dense model theorem with quasipolynomial dependencies, together with a quantitative generalized von Neumann lemma and Varnavides lower bounds (Teräväinen et al., 10 Mar 2026).

These formulations show that the term “transference principle” does not refer to a unique lemma. It denotes a family of mechanisms—Gowers-based, cut-norm-based, hypergraph-removal-based, and Fourier-analytic—that all serve the same structural function: approximate a sparse object by a dense bounded model and transfer configuration counts.

6. Extensions, variants, and alternatives

The principle rapidly generalized beyond one-dimensional arithmetic progressions in the primes. A multidimensional version proves that if $\nu$35, $\nu$36 is finite, and $\nu$37 has density at least $\nu$38, then for sufficiently large $\nu$39 the set $\nu$40 contains a constellation of shape $\nu$41. The proof uses the Green–Tao $\nu$42-trick, the linear forms condition for $\nu$43-tricked prime weights, a Varnavides-type averaging argument over homothetic grids, and the Furstenberg–Katznelson multidimensional Szemerédi theorem (Fox et al., 2013).

The same philosophy extends to nonclassical arithmetic settings. For coordinate rings of affine curves over finite fields, one chooses a polynomial subring $\nu$44 via Riemann–Roch, defines arithmetic progressions as $\nu$45-homothetic copies, constructs a pseudorandom measure from Goldston–Yıldırım weights and a $\nu$46-trick, and applies a relative Szemerédi theorem to prove that prime elements of $\nu$47 contain arbitrarily long arithmetic progressions (Kai, 2021).

Transference has also been adapted to finite-complexity affine-linear systems beyond translation-invariant progressions. One such theorem treats arbitrary finite-complexity affine-linear configurations, requiring only a linear forms condition on the majorant and density on higher-order Bohr sets, with applications to Chen primes, bounded-gap primes, and primes of the form $\nu$48 (Bienvenu et al., 2021). Other implementations address squares and $\nu$49th powers of primes through Fourier-analytic dense models and restriction estimates rather than hypergraph counting (Browning et al., 2015, Chow, 2016). Further adaptations prove arbitrarily long progressions inside primes of the form $\nu$50 and inside Piatetski–Shapiro primes by combining transference with specialized sieve weights and oscillatory estimates (Sun et al., 2017, Li et al., 2019).

There are also explicit surveys of Fourier-analytic dense model lemmas, including bounded, $\nu$51, and $\nu$52 approximants with Fourier-side control (Prendiville, 2015). More recently, arithmetic regularity has been proposed as an alternative to transference in settings where no obvious dense model is forthcoming; in that framework, the counting problem is decomposed into real, $\nu$53-adic, and combinatorial factors rather than transferred through a majorant–model pair (Chow et al., 1 Jun 2026).

The enduring significance of the Green–Tao transference principle lies in this portability. Once a sparse arithmetic set can be majorized by a measure with the correct linear-forms behavior, dense additive-combinatorial theorems become available far beyond their original domain.