QUANTITATIVE BOUNDS FOR THE U 4 -INVERSE THEOREM OVER LOW

CHARACTERISTIC FINITE FIELDS

JONATHAN TIDOR
arXiv:2109.13108v1 [math.CO] 27 Sep 2021

Abstract. This paper gives the first quantitative bounds for the inverse theorem for the Gowers
U 4 -norm over Fn
p when p = 2, 3. We build upon earlier work of Gowers and Milićević who solved the
corresponding problem for p ≥ 5. Our proof has two main steps: symmetrization and integration
of low-characteristic trilinear forms. We are able to solve the integration problem for all k-linear
forms, but the symmetrization problem we are only able to solve for trilinear forms. We pose
several open problems about symmetrization of low-characteristic k-linear forms whose resolution,
combined with recent work of Gowers and Milićević, would give quantitative bounds for the inverse
theorem for the Gowers U k+1 -norm over Fn p for all k, p.

1. Introduction
A central problem in additive combinatorics is to understand the inverse theory of the Gowers
uniformity norms. Given a 1-bounded function f : G → C with large Gowers uniformity norm
where G is a finite abelian group, the goal is to show that f must correlate with some structured
object.
This problem has primarily received attention in the cases that G is a cyclic group of prime
order or a finite field vector space. For G = Z/N Z the structured objects require the theory of
nilsequences to discribe, as proved by Green, Tao, and Ziegler [9, 10]. For finite field vector spaces
the situation is known to be somewhat simpler. For p ≥ k, a 1-bounded function f : Fnp → C
has large U k+1 -norm if and only if it correlates with a (classical) polynomial phase function, i.e.,
e2πiP (x)/p for a polynomial P : Fnp → Fp of degree at most k, as shown by Bergelson, Tao, and
Ziegler [2, 16] (see also [3] for a discussion of the p = k case). For p < k, one needs the theory of
non-classical polynomials to describe the structured objects [17].
The original proofs of these inverse theorems give extremely bad or even ineffective quantitative
bounds. Recently Manners proved quantitative bounds for the U k+1 -inverse theorem over Z/N Z
[14] and Gowers and Milićević proved quantitative bounds for the U k+1 -inverse theorem over Fnp
when p > k [6, 5].
The goal of this paper is to build upon the work of Gowers and Milićević to give quantitative
bounds for the U 4 -inverse theorem over Fnp in the low-characteristic regime p = 2, 3. Our proof has
two main steps: symmetrization and integration of low-characteristic trilinear forms. We are able
to solve the integration problem for all k-linear forms, but the symmetrization problem we are only
able to solve for trilinear forms.

1.1. Statement of main result. In this paper we always use V to denote a finite-dimensional
Fp -vector space and we write ω = e2πi/p . Given a function f : V → C and a shift h ∈ V , we write
∂h f (x) = f (x + h)f (x) for the multiplicative derivative.
Definition 1.1. Given a function f : V → C and d ≥ 2, the Gowers uniformity norm kf kU d is
defined by
d
kf k2U d = Ex,h1,...,hd ∈V (∂h1 · · · ∂hd f )(x).

The author was supported by NSF Graduate Research Fellowship Program DGE-1745302.
1
2 JONATHAN TIDOR

This definition has many useful properties, including being a well-defined norm for all d ≥ 2. See
[17, Lemma B.1] for more properties and references about the Gowers uniformity norms.
Consider a function f : V → G where G is an abelian group. Given a shift h ∈ V , we write
∆h f (x) = f (x + h) − f (x) for the additive derivative. It follows from the definition of the Gowers
uniformity norm that a 1-bounded function f : V → C satisfies kf kU d ≤ 1 with equality if and only
if f = e2πiP where P : V → R/Z satisfies ∆h1 ∆h2 · · · ∆hd P (x) = 0.
Definition 1.2. A non-classical polynomial of degree at most k is a map P : V → R/Z that
satisfies
(∆h1 · · · ∆hk+1 P )(x) = 0
for all h1 , . . . , hk+1 , x ∈ V . We write Poly 6k (V → R/Z) for the set of non-classical polynomials of
degree at most k.
A classical polynomial is a map P : V → Fp satisfying the same definition. Thus we use
Poly6k (V → Fp ) to denote the set of classical polynomials of degree at most k. In this paper
we commonly abuse notation to identify Fp with {0, 1/p, . . . , (p − 1)/p} ⊂ R/Z when convenient.
In this way we consider Poly 6k (V → Fp ) to be a subset of Poly6k (V → R/Z).
Theorem 1.3. Fix a prime p and δ > 0. There exists
    
1/ǫ = exp exp exp Op log(1/δ)O(1)
such that the following holds. Let V be a finite-dimensional Fp -vector space. Given a function
f : V → C satisfying kf k∞ ≤ 1 and kf kU 4 > δ, there exists a non-classical cubic polynomial
P ∈ Poly63 (V → R/Z) such that
 
Ex∈V f (x)e−2πiP (x)  ≥ ǫ.
 

Furthermore, if p ≥ 3, the polynomial can be taken to be classical.

This theorem was proved with ineffective bounds by Tao and Ziegler [17, Theorem 1.10] and
with effective bounds for p ≥ 5 by Gowers and Milićević [5, Theorem 7]. The bounds we prove are
of the same shape as Gowers and Milićević’s result.
1.2. Proof strategy. We start by summarizing the Gowers-Milićević proof of the U 4 -inverse the-
orem for p ≥ 5 and then explain where the difficulties arise when p < 5.
We are given a function f : Fnp → C which is 1-bounded and satisfies kf kU 4 > δ. By expanding
the definition of the U 4 -norm and applying the U 2 -inverse theorem, this implies
δ16 < kf k16 4 \ 2
U 4 = Ea,b,x k∂a ∂b f (x)kU 2 ≤ Ea,b k∂a ∂b f k∞ .
By an application of Markov’s inequality, this implies that there is a fairly dense set A ⊂ V 2 and
a function φ : A → V such that |∂\
a ∂b f (φ(a, b))| is large for each (a, b) ∈ A.
One can show that the function φ has some weak bilinearity properties. The main bulk of the
work done by Gowers and Milićević is to show that this local bilinearity can be upgraded to some
global structure. Namely, they prove that φ agrees with a biaffine map Φ : V 2 → V on a fairly
dense set. This immediately implies that there is a triaffine form T : V 3 → Fp that satisfies
 
Ea,b,c,x∂a ∂b ∂c f (x)ω T (a,b,c)  ≥ η. (1.1)
 

This part of the proof has no assumptions on the characteristic p.

To finish the proof, we would like to find a cubic polynomial P such that kf ω P kU 3 is large. From
this point we would be able to conclude by a single application of the U 3 -inverse theorem. We
expand
kf ω P k8U 3 = Ea,b,c,x ∂a ∂b ∂c f (x)ω ∆a ∆b ∆c P (x) .
 QUANTITATIVE U 4 -INVERSE THEOREM OVER LOW CHARACTERISTIC FINITE FIELDS 3

Thus the object ∆a ∆b ∆c P (x) naturally appears. We call this object the total derivative of P . The
total derivative of P clearly does not depend on x, and is symmetric and trilinear in a, b, c.
Thus our next step is to massage Eq. (1.1) to turn the triaffine form T into a symmetric trilinear
form. For simplicity of exposition, let us assume that T is a trilinear. (The parts of T that are not
trilinear can often be removed by an appropriate application of the Cauchy-Schwarz inequality.)
Symmetrization: We start with an argument based on an idea of Green and Tao. By several
applications of the Cauchy-Schwarz inequality, Eq. (1.1) implies that T (a, b, c) is “close” to T (b, a, c),
more precisely, we can prove that the map (a, b, c) 7→ T (a, b, c) − T (b, a, c) has bounded rank, and
similarly for all other permutations of a, b, c.1
To conclude the symmetrization step, we define
1
S(a, b, c) = (T (a, b, c) + T (a, c, b) + T (b, a, c) + T (b, c, a) + T (c, a, b) + T (c, b, a)) . (1.2)
6
Clearly S is a symmetric trilinear form. Furthermore, by the Green-Tao argument, we know that
rank(T − S) ≪η 1.
Integration: We now integrate the symmetric trilinear form S to a cubic polynomial P . Namely,
define
1
P (x) = S(x, x, x). (1.3)
6
One can easily check that ∆a ∆b ∆c P = S(a, b, c).
Starting from Eq. (1.1), with some involved but not particularly interesting additional inequali-
ties, we can conclude that kf ω P kU 3 ≫η 1. Applying the U 3 -inverse theorem lets us complete the
proof.
1.3. Low characteristic. Now we explain the difficulties that appear in characteristic p < 5.
Fortunately the bulk of the proof, up to Eq. (1.1), works in arbitrary characteristic. The Green-
Tao argument also goes through, showing that T is close to each of its permutations. However,
the remainder of the symmetrization step and the integration step, which were essentially trivial
in high characteristic, fail badly when p < 5. In particular, Eq. (1.2) and Eq. (1.3) both involve
dividing by 6.
For p ≥ 5 we have the convenient classification that S satisfies S(a, b, c) = ∆a ∆b ∆c P for some
cubic polynomial P if and only if S is a symmetric trilinear form. However, this is no longer the
case in low characteristic.
Following Tao and Ziegler, we define a classical symmetric trilinear form (or CSM for short) to be
a trilinear form S : V 3 → Fp that satisfies S(a, b, c) = ∆a ∆b ∆c P for some classical cubic polynomial
P . As part of their proof of the inverse theorem Tao and Ziegler classified CSMs. In the same
vein, we define a non-classical symmetric trilinear form (or nCSM for short) to be S : V 3 → Fp that
satisfies S(a, b, c) = ∆a ∆b ∆c P for some non-classical cubic polynomial P . One problem we solve
in this paper is the classification of non-classical symmetric multilinear forms.
In particular, we show that if S is a symmetric trilinear form and also has some further symmetries
which will be defined in Section 3, then we can integrate S to a non-classical cubic polynomial P
(i.e., ∆a ∆b ∆c P = S(a, b, c)). We are able to classify non-classical symmetric k-linear forms for all
k, which solves the integration problem necessary for the U k+1 -inverse theorem for all k.
For the symmetrization problem, we have a trilinear form T which we know is close in rank to
all of its permutations. We wish to find, for p = 2, an nCSM S such that rank(T − S) ≪ 1 and,
for p = 3, a CSM S such that rank(T − S) ≪ 1.
Even the problem of finding a symmetric multilinear S that is close to T is quite hard and we
do not know how to solve it for general k. When k = 3, the hypothesis that T is close to all of
its permutations implies that there is a subspace U ≤ V of bounded codimension so that T |U is
1See Section 2 for the definition of tensor rank that we use in this paper.
4 JONATHAN TIDOR

symmetric. This lets us easily find a symmetric trilinear S ′ such that rank(T − S ′ ) ≪ 1. For p = 3,
turning S ′ into a CSM is easy. For p = 2, we need to turn S ′ into an nCSM which takes some more
work, but we are able to do.
Structure of paper. We give the necessary preliminaries in Section 2. We solve the integration
problem for all k in Section 3 and the symmetrization problem for k = 3 in Section 4. We combine
these tools to prove our main result in Section 5. In Section 6 we give some conjectures.
Acknowledgements. The author thanks Aaron Berger for many helpful discussions.

2. Preliminaries
We gave a local definition of non-classical polynomials in Section 1. There is an equivalent
global definition known. (Also see [17, Lemma 1.7] for some more basic facts about non-classical
polynomials.)
Lemma 2.1 ([17, Lemma 1.7(iii)]). P : Fnp → R/Z is a non-classical polynomial of degree at most
k if and only if it can be expressed in the form
X ci1 ,...,in ,j |x1 |i1 · · · |xn |in
P (x1 , . . . , xn ) = α + (mod 1),
0≤i ,...,i <p, j≥0:
pj+1
1 n
0<i1 +···+in ≤k−j(p−1)

for some α ∈ R/Z and coefficients ci1 ,...,in ,j ∈ {0, . . . , p − 1} where | · | is the standard map Fp →
{0, . . . , p − 1}. Furthermore, this representation is unique.
Definition 2.2. Given a k-linear form T : V k → Fp we use two different notions for the rank of T .
The analytic rank of T , denoted arank(T ), is defined by the equation
h i
p− arank(T ) = Eh1 ,...,hk ∈V ω T (h1 ,...,hk ) .

The partition rank of T , denote prank(T ), is the length of the shortest decomposition of T as
a sum of partition rank 1 forms. We define T : V k → Fp to have partition rank 1 if we can write
T (h1 , . . . , hk ) = R(hI )S(h[k]\I ) where R : V I → Fp and S : V [k]\I → Fp are multilinear forms and
∅=6 I ( [k].
These definitions of rank enjoy many useful properties. One which will be important to us is the
subadditivity of rank. Obviously prank(S + T ) ≤ prank(S) + prank(T ). For analytic rank, we also
have arank(S + T ) ≤ arank(S) + arank(T ), but the proof is more subtle [12, Theorem 1.5].
For k = 2, the partition rank and analytic rank coincide and both are equal to the usual notion
of matrix rank. For general k we have the following bounds.
Theorem 2.3. There are constants Cp,k and Dk such that for a k-linear form T : V k → Fp we
have the bounds
arank(T ) ≤ prank(T ) ≤ Cp,k arank(T )Dk .
The lower bound is due to Lovett [12] and the upper bound independently by Milićević [15] and
Janzer [11]. Note that a recent result of Cohen and Moshkovitz and independently Adiprasito,
Kazhdan, and Ziegler improves the upper bound to linear when k = 3 and p > 2 [4, 1].

3. Non-classical symmetric multilinear forms

Definition 3.1. A non-classical symmetric multilinear form is a map T : V k → Fp that is:
• (multilinear) fixing all the variables but one, the map hi 7→ T (h1 , . . . , hk ) is a linear map;
• (symmetric) T (h1 , . . . , hk ) is invariant under permutations of the k variables h1 , . . . , hk ;
 QUANTITATIVE U 4 -INVERSE THEOREM OVER LOW CHARACTERISTIC FINITE FIELDS 5

• (non-classical) T (h1 , . . . , h1 , h2 , . . . , hk−p+1 ) = T (h1 , h2 , . . . , h2 , h3 , . . . , hk−p+1 ) where the

variable h1 appears p times in the first expression and h2 appears p times in the second
expression.
We write nCSMk (V ) for the space of non-classical symmetric multilinear forms T : V k → Fp .
Non-classical symmetric multilinear forms were implicitly introduced by Tao and Ziegler in their
proof of the inverse theorem [17]. There they study the related notion of classical symmetric
multilinear forms which satisfy the more restrictive condition that T (h1 , . . . , h1 , h2 , . . . , hk−p+1 ) = 0.
Definition 3.2. For P ∈ Poly6k (V → R/Z), define the total derivative of P to be the map
dk P : V k → Fp defined by
dk P (h1 , . . . , hk ) = (∆h1 · · · ∆hk P )(0).
Note that the right-hand side of this equation lies in {0, 1/p, . . . , (p−1)/p} ⊂ R/Z which we identify
with Fp . Also since P is a non-classical polynomial of degree at most k we have dk P (h1 , . . . , hk ) =
(∆h1 · · · ∆hk P )(x) for all h1 , . . . , hk , x ∈ V .
Tao and Ziegler prove that the total derivatives of classical polynomials are exactly CSMs.
Proposition 3.3 ([17, Lemma 4.5]). For k ≥ 1 and a classical polynomial P ∈ Poly6k (V → Fp ),
the total derivative dk P lies in CSMk (V ). Furthermore this map is surjective. In other words, we
have the short exact sequence
dk
0 → Poly 6k−1 (V → Fp ) → Poly 6k (V → Fp ) −→ CSMk (V ) → 0.
They also show that the total derivative of a non-classical polynomial is an nCSM.
Proposition 3.4 ([17]). For P ∈ Poly6k (V → R/Z), the total derivative dk P lies in nCSMk (V ).
Our integration result gives the implication in the opposite implication.
Proposition 3.5 (Integrating nCSMs). For k ≥ 1 and T ∈ nCSMk (V ), there exists a non-classical
polynomial P ∈ Poly6k (V → R/Z) such that dk P = T . In other words, we have the short exact
sequence
dk
0 → Poly6k−1 (V → R/Z) → Poly6k (V → R/Z) −→ nCSMk (V ) → 0.
Proof. First note that ker(dk ) = Poly 6k−1 (V → R/Z) since ∆h1 · · · ∆hk P is identically zero for all
h1 , . . . , hk if and only if P is a non-classical polynomial of degree at most k − 1.
We show that dk is surjective in an inexplicit manner. First define Poly 6k (V → R/Z) to be
Poly6k (V → R/Z) under the equivalence relation [P ] = [P + α] for α ∈ R/Z. Note that for k ≥ 1,
the total derivative dk acts on this quotient space. We will prove that
|Poly 6k (V → R/Z)|
= | nCSMk (V )|. (3.1)
|Poly 6k−1 (V → R/Z)|
This suffices to complete the proof.
Choose an isomorphism V ∼
′
= Fnp . By Lemma 2.1, we see that |Poly6k (V → R/Z)| = pCk
where Ck′ is defined to be the number of tuples (i1 , . . . , in , j) where i1 , . . . , in ∈ {0, . . . , p − 1} and
′ ′
j ≥ 0 satisfy 0 < i1 + · · · + in ≤ k − j(p − 1). Then the left-hand side of Eq. (3.1) is pCk −Ck−1 .
Define Ck = Ck′ − Ck−1 ′ . We can easily see that Ck is the number of tuples (i1 , . . . , in , j) where
i1 , . . . , in ∈ {0, . . . , p − 1} and 0 ≤ j ≤ (k − 1)/(p − 1) satisfy i1 + · · · + in = k − j(p − 1).
Now we compute | nCSMk (V )|. Clearly any multilinear form T : V k → Fp can be expressed
uniquely as X
T (x1 , . . . , xk ) = cj1 ,...,jk x1,j1 · · · xk,jk
j1 ,...,jk ∈[n]
6 JONATHAN TIDOR

where cj1 ,...,jk ∈ Fp . For a tuple j = (j1 , . . . , jk ), define i(j) = (i1 , . . . , in ) such that iℓ is the
number of j1 , . . . , jk which are equal to ℓ. With this notation, we can say that a multilinear form
T is symmetric if and only if cj1 = cj2 whenever i(j1 ) = i(j2 ).
Finally define i′ (j) = (i′1 , . . . , i′n ) where i′ℓ = 0 if iℓ = 0 and otherwise i′ℓ ∈ {1, . . . , p − 1}
satisfies i′ℓ ≡ iℓ (mod p − 1). We can see that a multilinear form T is an nCSM if and only if
cj1 = cj2 whenever i′ (j1 ) = i′ (j2 ). Thus | nCSMk (V )| = pDk where Dk is the number of tuples
i′ ∈ {0, . . . , p − 1}n such that there exists j ∈ [n]k such that i′ = i′ (j). Clearly the only constraint
on such a tuple i′ ∈ {0, . . . , p − 1}n is that s = i′1 + · · · i′n satisfies 0 < s ≤ n and s ≡ n (mod p − 1).
Thus we see that Ck = Dk , proving the desired result. 

4. Symmetrization in low characteristic

For a k-linear form T : V k → Fp and a permutation π ∈ Sk , write Tπ : V k → Fp for
Tπ (x1 , . . . , xk ) = T (xπ(1) , . . . , xπ(k) ).
We have one tool that lets us show that forms are close to symmetric. The following result is
based on an idea of Green and Tao.
Lemma 4.1 (cf. [7, Lemma 2.8]). Let A : V 2 → Fp be a bilinear form. Let b1 , b2 , b3 : V → C be
arbitrary 1-bounded functions. If
 
A(u,v) 
Eu,v∈V b1 (u)b2 (v)b3 (u + v)ω  ≥ δ,


then
p− arank(A−A(12) ) = Eu,v ω A(u,v)−A(v,u) ≥ δ8 .
Proof. By an application of the Cauchy-Schwarz inequality and the 1-boundedness of b2 , we have
′
δ2 ≤ Eu,u′ ,v∈V b1 (u)b1 (u′ )b3 (u + v)b3 (u′ + v)ω A(u−u ,v) .
We make the change of variables u′′ = u + u′ + v. Rearranging, we obtain
′ ′′ ′
δ2 ≤ Eu,u′ ,u′′ b1 (u)b1 (u′ )b3 (u′′ − u′ )b3 (u′′ − u)ω A(u−u ,u −u−u )
 ′′
 ′ ′′ ′
 ′ ′
= Eu,u′ ,u′′ b1 (u)b3 (u′′ − u)ω A(u,u −u) b1 (u′ )b3 (u′′ − u′ )ω A(−u ,u −u ) ω A(u ,u)−A(u,u ) .

Write B = A − A(12) . By first averaging over u′′ and then choosing b′1 , b′2 : V → C appropriate
1-bounded functions, the above inequality becomes
 ′ 

δ2 ≤ Eu,u′ b′1 (u)b′2 (u′ )ω B(u,u )  .


By two more applications of the Cauchy-Schwarz inequality, we conclude

′ ′ ′
δ8 ≤ Eu,v,u′ ,v′ ω B(u−v,u −v ) = Eu,u′ ω B(u,u ) = p− arank(B) . 
Corollary 4.2. Let T : V 3 → Fp be a trilinear form. Suppose there are 1-bounded functions
b1 , . . . , b7 : V → C such that
 
Ex,y,z∈V b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)(−1)T (x,y,z)  ≥ δ.
 

Then we have arank(T − Tπ ) ≤ 16 logp (1/δ) for all π ∈ S3 .

Proof. Suppose π = (12). By the triangle inequality, we see that there exists a function δ : V → [0, 1]
and 1-bounded functions b1,z , b2,z , b3,z : V → C such that
 
Ex,y b1,z (x)b2,z (y)b3,z (x + y)ω T (x,y,z) ≥ δ(z)
 

for all z ∈ V and additionally Ez∈V δ(z) ≥ δ.

 QUANTITATIVE U 4 -INVERSE THEOREM OVER LOW CHARACTERISTIC FINITE FIELDS 7

By Lemma 4.1 for all z ∈ V ,

Ex,y ω T (x,y,z)−T (y,x,z) ≥ δ(z)8 .
Averaging over z and applying convexity,
p− arank(T −T(12) ) = Ex,y,z ω T (x,y,z)−T (y,x,z) ≥ Ez δ(z)8 ≥ δ8 ,
so arank(T − T (12)) ≤ 8 logp (1/δ). The same holds for all transpositions. Since every permutation
in S3 can be written as the product over two transpositions, we conclude by the subadditivity of
analytic rank. 
Proposition 4.3. Let T : V 3 → Fp be a trilinear form. Suppose that T is close to symmetric in
the sense that prank(T − Tπ ) ≤ r for all permutations π ∈ S3 . Then there exists a subspace U ≤ V
with codim U ≤ 5r such that T |U : U 3 → Fp is a symmetric trilinear form.
Proof. By definition, there exists a partition rank decomposition of each T − Tπ (for π ∈ S3 \ {id})
as the sum of at most r terms, each of which is the product of a linear form and a bilinear form.
Let U be the subspace of V where all the linear forms vanish. Since there are at most 5r linear
forms in this decomposition, we have the desired bound on the codimension of U . Furthermore,
since (T − Tπ )|U is identically zero for each π ∈ S3 , we see that T |U is symmetric. 
Remark. This result immediately implies that there exists a symmetric trilinear form S : V 3 → Fp
with prank(T − S) ≤ 15r. To see this, pick an arbitrary decomposition V = U ⊕ W . Then define
S(u1 ⊕ w1 , u2 ⊕ w2 , u3 ⊕ w3 ) = T |U (u1 , u2 , u3 ). We have the bound prank(T − S) ≤ 15r since T − S
can be written as a linear combination of linear and bilinear forms where the linear forms are given
by dotting with a basis element of W and the bilinear forms are arbitrary.
Proposition 4.3 easily handles the initial symmetrization step. To complete the symmetrization,
we need to find some R ∈ nCSM3 (V ) such that prank(T − R) is small. For p ≥ 5 this is trivial,
since every symmetric trilinear form is an nCSM and actually a CSM.
For p = 3 every symmetric trilinear form is also an nCSM though not necessarily a CSM. We
note however, that for p = 3 every nCSM is close to a CSM. This allows us to prove the U 4 -inverse
theorem over Fn3 with classical polynomials.
Proposition 4.4. Let T : V 3 → F3 be a symmetric trilinear form. Then there exists a codimension
1 subspace U ≤ V such that T |U ∈ CSM3 (U ).
Proof. The only additional condition that T needs to satisfy to be a CSM is T (x, x, x) = 0. Thus
if U is any subspace of V such that T (x, x, x) = 0 for all x ∈ U , we would have T |U ∈ CSM3 (U ).
Consider the map x 7→ T (x, x, x) ∈ F3 . Because T is a symmetric trilinear form, this map
is linear. Indeed, T (x + y, x + y, x + y) = T (x, x, x) + 3T (x, x, y) + 3T (x, y, y) + T (y, y, y) =
T (x, x, x) + T (y, y, y). Defining U to be the codimension 1 subspace on which T (x, x, x) vanishes,
we have the desired conclusion. 
Corollary 4.5. Fix p ≥ 3. Let T : V 3 → Fp be a trilinear form that satisfies prank(T − Tπ ) ≤ r
for all permutations π ∈ S. Then there exists S ∈ CSM3 (V ) such that prank(T − S) ≤ 15r + 3.
For p = 2 the situation is more complicated. It is known, due to examples of Green and Tao [8]
and Lovett, Meshulam, and Samorodnitsky [13] that classical polynomials are not sufficient for the
U 4 -inverse theorem over F2 . This means that we will not be able to find a CSM that is close to T .
However the following argument shows that T is close to an nCSM.
Proposition 4.6. Let T : V 3 → F2 be a symmetric trilinear form. Suppose there are 1-bounded
functions b1 , . . . , b7 : V → C such that
 
Ex,y,z∈V b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)(−1)T (x,y,z)  ≥ δ.
 

Then there exists a subspace U ≤ V satisfying codim U ≤ 8 log2 (1/δ) such that T |U ∈ nCSM3 (U ).
8 JONATHAN TIDOR

Proof. Make the change of variables z = x + w. Then the hypothesis becomes

 
T (x,x,y)+T (x,y,w) 
δ ≤ Ex,y,w∈V b1 (x)b2 (y)b3 (x + w)b4 (x + y)b5 (w)b6 (x + y + w)b7 (y + w)(−1)


 
T (x,x,y)+T (x,y,w) 
= Ew b5 (w)Ex,y (b1 (x)b3 (x + w))(b2 (y)b7 (y + w))(b4 (x + y)b6 (x + y + w))(−1) .


By averaging (and the 1-boundedness of b5 ), there exists w0 ∈ V such that

 
δ ≤ Ex,y (b1 (x)b3 (x + w0 ))(b2 (y)b7 (y + w0 ))(b4 (x + y)b6 (x + y + w0 ))(−1)T (x,x,y)+T (x,y,w0 )  .
 

Consider the map A : V 2 → F2 defined by A(x, y) = T (x, x, y) + T (x, y, w0 ). We claim that A is

bilinear. Using the symmetry and trilinearity of T ,

A(x + x′ , y) = T (x + x′ , x + x′ , y) + T (x + x′ , y, w0 )
= T (x, x, y) + 2T (x, x′ , y) + T (x′ , x′ , y) + T (x, y, w0 ) + T (x′ , y, w0 )
= A(x, y) + A(x′ , y).

Similarly,

A(x, y + y ′ ) = T (x, x, y + y ′ ) + T (x, y + y ′ , w0 )

= T (x, x, y) + T (x, x, y ′ ) + T (x, y, w0 ) + T (x, y ′ , w0 )
= A(x, y) + A(x, y ′ ).

Since A is bilinear, we can apply Lemma 4.1 to conclude that

δ8 ≤ Ex,y (−1)A(x,y)−A(y,x) .

Note that

A(x, y) − A(y, x) = T (x, x, y) + T (x, y, w0 ) − T (y, y, x) − T (y, x, w0 ) = T (x, x, y) − T (x, y, y)

by the symmetry of T .
Define B : V 2 → F2 by B(x, y) = T (x, x, y) − T (x, y, y). Note that B is bilinear. (This can be
seen since B(x, y) = A(x, y) − A(y, x) or by direct computation.) We have shown that arank B ≤
8 log2 (1/δ). Let U be the nullspace of B. This satisfies codim U = prank B = arank B ≤ 8 log2 (1/δ).
Furthermore B(x, y) = 0 for all x, y ∈ U , which implies that T |U is an nCSM. 

This essentially implies the desired symmetrization result. To put the result in the form we
need, we use the following application of the Cauchy-Schwarz inequality which will also appear
again later in the paper.

Lemma 4.7. Let φ : V 3 → Fp be a triaffine form. Suppose there are 1-bounded functions b1 , . . . , b7 : V →
C such that
 
φ(x,y,z) 
b (x)b (y)b (z)b (x + y)b (x + z)b (y + z)b (x + y + z)ω  ≥ δ.

Ex,y,z∈V 1 2 3 4 5 6 7

Let T : V 3 → Fp be the trilinear part of φ. Then there exist 1-bounded functions b′1 , . . . , b′7 : V → C
such that
 
Ex,y,z∈V b′1 (x)b′2 (y)b′3 (z)b′4 (x + y)b′5 (x + z)b′6 (y + z)b′7 (x + y + z)ω T (x,y,z)  ≥ δ8 .
 
 QUANTITATIVE U 4 -INVERSE THEOREM OVER LOW CHARACTERISTIC FINITE FIELDS 9

Proof. This result follows from three applications of the Cauchy-Schwarz inequality, one in each
variable. We have
 2
δ2 ≤ Ey,z b2 (y)b3 (z)b6 (y + z)Ex b1 (x)b4 (x + y)b5 (x + z)b7 (x + y + z)ω φ(x,y,z) 
 
 
≤ Ey,z |b2 (y)b3 (z)b6 (y + z)|2

· Ex,x′ ,y,z b1 (x)b1 (x′ )b4 (x + y)b4 (x′ + y)b5 (x + z)b5 (x′ + z)
′

· b7 (x + y + z)b7 (x′ + y + z)ω φ(x−x ,y,z) .

The 1-boundedness of b2 , b3 , b6 lets us remove the first term.

Repeating two more times gives
 ′ ′ ′ 

δ8 ≤ Ex,x′ ,y,y′ ,z,z ′ b(x, x′ , y, y ′ , z, z ′ )ω φ(x−x ,y−y ,z−z ) 


where b is a product of the b7 ’s evaluated at sums of the six variables. Note that φ(x − x′ , y − y ′ , z −
z ′ ) = T (x − x′ , y − y ′ , z − z ′ ). Then by averaging, we can find x0 , y0 , z0 such that
 
δ8 ≤ Ex,y,z b(x + x0 , x0 , y + y0 , y0 , z + z0 , z0 )ω T (x,y,z)  .
 

Tracing through the proof, we can see that b(x + x0 , x0 , y + y0 , y0 , z + z0 , z0 ) has the desired
factorization. 
Corollary 4.8. Let T : V 3 → F2 be a trilinear form. Suppose there are 1-bounded functions
b1 , . . . , b7 : V → C such that
 
Ex,y,z∈V b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)(−1)T (x,y,z)  ≥ δ.
 

Then there exists S ∈ nCSM3 (V ) such that prank(T − S) ≤ 432 log 2 (1/δ).
Proof. By Corollary 4.2, we have that arank(T − Tπ ) ≤ 16 log 2 (1/δ) for all π ∈ S3 . Then applying
Proposition 4.3, there exists a subspace U ≤ V with codim U ≤ 80 log 2 (1/δ) such that T |U : U 3 →
F2 is a symmetric trilinear form.
By averaging, there exist cosets x0 + U, y0 + U, z0 + U such that
 
 
T (x,y,z)
 
δ ≤ Ex∈x0 +U b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)(−1) 
 y∈y0 +U


 z∈z0 +U 
= Ex,y,z∈U b′1 (x)b′2 (y)b′3 (z)b′4 (x + y)b′5 (x + z)b′6 (y + z)b′7 (x + y + z)(−1)T (x0 +x,y0+y,z0 +z)  .
 

Now by Lemma 4.7, this implies that

 
8 ′′ ′′ ′′ ′′ ′′ ′′ ′′ T (x,y,z) 
δ ≤ Ex,y,z∈U b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)(−1) .


We can now apply Proposition 4.6 to find a subspace W ≤ U such that codim W ≤ 144 log 2 (1/δ)
and such that T |W ∈ nCSM3 (W ). Pick an arbitrary decomposition V = W ⊕ W ′ and then define
S(w1 ⊕ w1′ , w2 ⊕ w2′ , w3 ⊕ w3′ ) = T |W (w1 , w2 , w3 ). Then S ∈ nCSM3 (V ) and clearly prank(T − S) ≤
432 log 2 (1/δ). 

5. Proof of main theorem

We now have the tools to prove the main theorem. We first record the result of Gowers and
Milićević that we require.
10 JONATHAN TIDOR

Theorem 5.1 ([6]). Fix a prime p and δ > 0. There exists

     −1
ǫ = exp exp exp Op log(1/δ)O(1)
such that the following holds. Let V be a finite-dimensional Fp -vector space. Given a function
f : V → C satisfying kf k∞ ≤ 1 and kf kU 4 > δ, there exists a triaffine form φ : V 3 → Fp such that
 
Ex,h1,h2 ,h3 ∈V (∂h1 ∂h2 ∂h3 f )(x)ω φ(h1 ,h2 ,h3 )  ≥ ǫ. (5.1)
 

Proof of Theorem 1.3. Applying Theorem 5.1, we find a triaffine form φ : V 3 → Fp such that
 
Ex,h1,h2 ,h3 ∈V (∂h1 ∂h2 ∂h3 f )(x)ω φ(h1 ,h2 ,h3)  ≥ ǫ.
 

Let T be the trilinear part of φ. First average over x and expand out the derivative to conclude
that there exist 1-bounded functions b1 , . . . , b7 : V → C such that
 
φ(x,y,z) 
Ex,y,z∈V b1 (x)b2 (y)b3 (z)b4 (x + y)b5 (x + z)b6 (y + z)b7 (x + y + z)ω  ≥ ǫ.


Then applying Lemma 4.7 and then Corollary 4.2, we see that arank(T − Tπ ) ≤ 16 log p (1/ǫ) for all
π ∈ S3 . By Theorem 2.3, we conclude that prank(T − Tπ ) ≤ poly logp (1/ǫ) for all π ∈ S3 .
Now for p ≥ 3, we can apply Corollary 4.5 to find S ∈ CSM3 (V ) such that prank(T − S) ≤
poly logp (1/ǫ). Then by Proposition 3.3, there exists a classical cubic polynomial P ∈ Poly63 (V →
Fp ) such that d3 P = S. For p = 2, we apply Corollary 4.8 to find S ∈ nCSM3 (V ) such that
prank(T − S) ≤ poly logp (1/ǫ). Then by Proposition 3.5, there exists a non-classical cubic polyno-
mial P ∈ Poly63 (V → R/Z) such that d3 P = S.
Now in either case we have a decomposition T − S = ri=1 γi where each γi : V 3 → Fp is the
P
product of a linear form in one of the variables with a bilinear form in the other two variables and
r = poly logp (1/ǫ). Write Γ : V 3 → F2r p for the list of 2r linear and bilinear forms that factor the γi ’s.
For example, if γ1 (x, y, z) = α(x)β(y, z), then we define Γ1 (x, y, z) = α(x) and Γ2 (x, y, z) = β(y, z).
Define g = f ω P . Note that (∂h1 ∂h2 ∂h3 g)(x) = (∂h1 ∂h2 ∂h3 f )(x)ω S(h1 ,h2 ,h3 ) . The inequality we
started with can now be written as
 Pr 
Ex,h1,h2 ,h3 b(h1 , h2 , h3 )ω i=1 γi (h1 ,h2 ,h3) (∂h1 ∂h2 ∂h3 g)(x) ≥ ǫ
 

where b(h1 , h2 , h3 ) is a product of linear and bilinear phase functions. We first remove the middle
term.
By averaging, there exists c ∈ F2r p such that

|Ex,h1 ,h2,h3 b(h1 , h2 , h3 )1(Γ(h1 , h2 , h3 ) = c)(∂h1 ∂h2 ∂h3 g)(x)| ≥ ǫp−2r .

Now by Fourier analysis, we can write
1(Γ(h1 , h2 , h3 ) = c) = Eξ∈F2r
p
ω (Γ(h1 ,h2 ,h3)−c)·ξ .
Plugging this expression into the previous inequality, by averaging again, there exists ξ0 ∈ F2r p such
that  
Ex,h1,h2 ,h3 b(h1 , h2 , h3 )ω (Γ(h1 ,h2 ,h3)−c)·ξ0 (∂h1 ∂h2 ∂h3 g)(x) ≥ ǫp−2r .
 

Since Γ is a list of linear and bilinear forms, the middle term, ω (Γ(h1 ,h2 ,h3)−c)·ξ0 , is the product of
linear and bilinear phase functions. Thus we conclude that
Ex,h ,h ,h b′ (h1 , h2 , h3 )(∂h ∂h ∂h g)(x) ≥ ǫp−2r ,
 
1 2 3 1 2 3

b′
where is the product of linear and bilinear phase functions.
Write
b′ (h1 , h2 , h3 ) = ω β1 (h2 ,h3 )+β2 (h1 ,h3 )+β3 (h1 ,h2 )+α1 (h1 )+α2 (h2 )+α3 (h3 ) .
 QUANTITATIVE U 4 -INVERSE THEOREM OVER LOW CHARACTERISTIC FINITE FIELDS 11

By averaging, there exist x, h1 ∈ V such that

Eh ,h b′ (h1 , h2 , h3 )(∂h ∂h ∂h g)(x) ≥ ǫp−2r .
 
2 3 1 2 3

Expanding this out, we see that this inequality is of the correct form to apply Lemma 4.1 which
then implies that arank(β1 − (β1 )(12) ) ≤ 8(2r + logp (1/ǫ)) ≤ 24r. The same is true of β2 , β3 .
The next step is to find β1′ , β2′ , β3′ ∈ nCSM2 (V ) such that prank(βi − βi′ ) ≤ 48r. First note that
for bilinear forms we have arank(βi ) = prank(βi ). Thus there is a subspace U ≤ V of codimension
at most 24r such that βi − (βi )(12) vanishes on U × U . This is the only condition we need to imply
that βi |U is an nCSM. Extending βi |U to βi′ ∈ nCSM2 (V ) arbitrarily, we have prank(βi − βi′ ) ≤ 48r.
Finally, by Proposition 3.3, there exist non-classical quadratic polynomials Qi ∈ Poly62 (V → R/Z)
such that d2 Qi = βi′ .
We repeat the same averaging and Fourier analysis argument from above to conclude that
 ′ ′ ′

Ex,h1,h2 ,h3 b′′ (h1 , h2 , h3 )ω β1 (h2 ,h3 )+β2 (h1 ,h3 )+β3 (h1 ,h2) (∂h1 ∂h2 ∂h3 g)(x) ≥ ǫp−300r ,
 

where b′′ (h1 , h2 , h3 ) is a product just of linear phase functions, say b′′ (h1 , h2 , h3 ) = ω L1 (h1 )+L2 (h2 )+L3 (h3 ) .
Now define
g000 = f ω P ,
g001 = f ω P +Q1 ,
g010 = f ω P +Q2 ,
g011 = f ω P +Q1+Q2 +L3 ,
g100 = f ω P +Q3 ,
g101 = f ω P +Q1+Q3 +L2 ,
g110 = f ω P +Q2+Q3 +L1 ,
g111 = f ω P +Q1+Q2 +Q3 +L1 +L2 +L3 .
One can verify that the inequality above can be rewritten as
|Ex,h1,h2 ,h3 g000 (x) g001 (x + h1 )g010 (x + h2 )g011 (x + h1 + h2 )g100 (x + h3 )

·g101 (x + h1 + h3 )g110 (x + h2 + h3 )g111 (x + h1 + h2 + h3 ) ≥ ǫp−300r .


The left-hand side is in the correct form to apply the Gowers-Cauchy-Schwarz inequality [17,
O(1)
Lemma B.1.(iv)], which implies that kf ω P kU 3 ≥ ǫp−300r ≥ p−O(logp (1/ǫ) ) . Finishing with a
single application of the U 3 -inverse theorem, we prove the desired result. 

6. Concluding remarks
The symmetrization results in this paper are very specific to trilinear forms. We pose the
following conjecture, which seems surprisingly difficult and may be of independent interest.
Conjecture 6.1. Let T : V k → Fp be a k-linear form. Suppose that T is close to symmetric in the
sense that prank(T − Tπ ) ≤ r for all permutations π ∈ Sk . Does there exists a symmetric k-linear
form S : V k → Fp such that prank(S − T ) ≤ r ′ where r ′ is a polynomial function of r?
NotePthat this conjecture is only interesting if p ≤ k, since for p > k, the symmetric form
1
S = k! π Tπ can easily be seen to satisfy prank(S − T ) ≤ k!r.
Combined with the recent work of Gowers and Milićević [5], Conjecture 6.1 is the missing in-
gredient which would suffice to give quantitative bounds on the U p+1 -inverse theorem. For the
U k+1 -inverse theorem with k > p there is additional symmetrization necessary to produce a non-
classical symmetric multilinear form from a symmetric form.
12 JONATHAN TIDOR

Conjecture 6.2. Let T : V k → Fp be a k-linear form. Suppose that T satisfies

 ! 
 
Y X
T (x1 ,...,xk ) 
 
Ex ,...,x
 1 k
bI x i ω ≥ǫ
 I⊆[k] i∈I 

for some 1-bounded functions bI : V → C. Does there exists S ∈ nCSMk (V ) such that prank(S −
T ) ≤ r ′ where r ′ is a polynomial function of log(1/ǫ)?
Combined with the work of Gowers and Milićević and this paper’s integration result, Proposi-
tion 3.5, this result would give quantitative bounds on the U k+1 -inverse theorem over Fnp for all
k, p.

References
[1] Karim Adiprasito, David Kazhdan, and Tamar Ziegler, On the Schmidt and analytic ranks for trilinear forms,
arXiv:2102.03659.
[2] Vitaly Bergelson, Terence Tao, and Tamar Ziegler, An inverse theorem for the uniformity seminorms associated
with the action of F∞p , Geom. Funct. Anal. 19 (2010), 1539–1596.
[3] Aaron Berger, Ashwin Sah, Mehtaab Sawhney, and Jonathan Tidor, Non-classical polynomials and the inverse
theorem, arXiv:2107.07495.
[4] Alex Cohen and Guy Moshkovitz, Structure vs. randomness for bilinear maps, Proceedings of the 53rd Annual
ACM Symposium on Theory of Computing (STOC 2021), ACM, New York, 2021, pp. 800–808.
[5] W. T. Gowers and L. Milićević, An inverse theorem for Freiman multi-homomorphisms, arXiv:2002.11667.
[6] W. T. Gowers and L. Milićević, A quantitative inverse theorem for the U 4 norm over finite fields,
arXiv:1712.00241.
[7] Ben Green, Montréal notes on quadratic Fourier analysis, Additive combinatorics, CRM Proc. Lecture Notes,
vol. 43, Amer. Math. Soc., Providence, RI, 2007, pp. 69–102.
[8] Ben Green and Terence Tao, The distribution of polynomials over finite fields, with applications to the Gowers
norms, Contrib. Discrete Math. 4 (2009), 1–36.
[9] Ben Green, Terence Tao, and Tamar Ziegler, An inverse theorem for the Gowers U 4 -norm, Glasg. Math. J. 53
(2011), 1–50.
[10] Ben Green, Terence Tao, and Tamar Ziegler, An inverse theorem for the Gowers U s+1 [N ]-norm, Ann. of Math.
(2) 176 (2012), 1231–1372.
[11] Oliver Janzer, Polynomial bound for the partition rank vs the analytic rank of tensors, Discrete Anal. (2020),
Paper No. 7, 18.
[12] Shachar Lovett, The analytic rank of tensors and its applications, Discrete Anal. (2019), Paper No. 7, 10.
[13] Shachar Lovett, Roy Meshulam, and Alex Samorodnitsky, Inverse conjecture for the Gowers norm is false,
Theory Comput. 7 (2011), 131–145.
[14] Freddie Manners, Quantitative bounds in the inverse theorem for the Gowers U s+1 -norms over cyclic groups,
arXiv:1811.00718.
[15] Luka Milićević, Polynomial bound for partition rank in terms of analytic rank, Geom. Funct. Anal. 29 (2019),
1503–1530.
[16] Terence Tao and Tamar Ziegler, The inverse conjecture for the Gowers norm over finite fields via the correspon-
dence principle, Anal. PDE 3 (2010), 1–20.
[17] Terence Tao and Tamar Ziegler, The inverse conjecture for the Gowers norm over finite fields in low character-
istic, Ann. Comb. 16 (2012), 121–188.

Massachusetts Institute of Technology, Cambridge, MA 02139, USA

Email address: [email protected]