A bilinear approach to the finite field restriction problem, II

We may enlarge \mathcal L to the family of all lines meeting A in at most C m^{1/2} points, since every ordered pair of distinct points determines a unique line and all terms are nonnegative. Constants in this proof may depend on C. Put \mu=\min(m,p) and \tau=\mu^{1/3}. Split this family as

where the lines in \mathcal L^- meet A in at most \tau points and those in \mathcal L^+ meet A in more than \tau points. Since

the elementary inequality (u+v)^2\le2u^2+2v^2 gives

We first bound the low-richness part. In a fixed direction, the lines of \mathcal L^- are parallel and disjoint. If such a line contains r\le\tau points of A, then it contributes r(r-1)\le\tau r ordered pairs. Consequently,

Moreover, every ordered pair of distinct points has a unique direction, so \sum_\theta E_{\theta,\mathcal L^-}(A)\le m^2. Therefore

It remains to estimate \mathcal L^+. For dyadic parameters j and s, let \mathcal C_{j,s} be the lines in \mathcal L^+ meeting A in between j and 2j points and whose direction contains between s and 2s such lines. The nonempty classes form a collection \mathscr C of cardinality J=O((1+\log m)^2). For each direction,

and Cauchy–Schwarz over the classes gives

Thus it is enough to prove the asserted bound uniformly for one class. More explicitly, if every class contributes at most H, then the right side of (?) is at most J^2H. Since J=O((1+\log m)^2), this costs at most O((1+\log m)^4), which is absorbed into p^\varepsilon because m\le p^2.

Fix a nonempty class \mathcal C=\mathcal C_{j,s}, and let t be the number of directions it spans. Then j\gtrsim\tau, the class contains \sim s lines in each of these directions and \sim ts lines in all, and each line contains \sim j points of A. For a fixed direction the class contributes O(sj^2) ordered pairs, and hence

We will use three constraints repeatedly. There are p+1 directions, so t\lesssim p; the lines of one direction are parallel and disjoint, so sj\lesssim m; and the hypothesis gives j\lesssim m^{1/2}.

In this class each of the \sim ts lines carries at least j points, so

Whichever incidence bound applies, at least one term on its right side is \gtrsim jts. Suppose first that the class spans few directions, t\le c_0p^2/m, so that (3.1) applies. If the term t^{2/9}(ts)^{5/9}m^{7/9} in (3.1) is \gtrsim jts, then

At the same time, Vinh’s estimate (3.2) applies to this same class. If the term mL/p\sim mts/p in (3.2) is \gtrsim jts, then j\lesssim m/p, and using t\lesssim p and sj\lesssim m gives

If the term (pmL)^{1/2}\sim(pmts)^{1/2} in (3.2) is \gtrsim jts, then ts\lesssim pm/j^2, and hence, using sj\lesssim m,

Consequently, in the case where the term t^{2/9}(ts)^{5/9}m^{7/9} in (3.1) is \gtrsim jts, we have

If m\le p, then j\gtrsim\mu^{1/3}=m^{1/3} and the minimum in (?) is \lesssim m^{10/3}=m^3\mu^{1/3}. If m\ge p, put j_0=mp^{-2/3}. Since j_0\ge p^{1/3}=\mu^{1/3}, either j\le j_0, in which case pm^2j\le m^3p^{1/3}, or j\ge j_0, in which case m^{7/2}j^{-1/2}\le m^3p^{1/3}. Thus this case contributes at most a constant multiple of

We next handle the other terms in (3.1). If the term t^{1/2}(ts)^{1/2}m^{1/2} is \gtrsim jts, then s\lesssim m/j^2, so ts^2j^4\lesssim tm^2\lesssim pm^2. If the term (ts)^{2/3}m^{2/3} is \gtrsim jts, then ts\lesssim m^2j^{-3}, so ts^2j^4\lesssim m^2(sj)\lesssim m^3\le m^3\mu^{1/3}. If the term m is \gtrsim jts, then ts^2j^4=(jts)(sj^3)\lesssim m^3\le m^3\mu^{1/3}, using sj\lesssim m and j\lesssim m^{1/2}. Finally, the L term in (3.1) satisfies L\sim ts for the present class. If this term is \gtrsim jts, then j\lesssim1; using sj\lesssim m and t\lesssim p gives ts^2j^4\lesssim ts^2\lesssim t(m/j)^2\lesssim pm^2. Thus each class with few directions satisfies

Suppose instead that the class spans many directions, t>c_0p^2/m, which is outside the stated range of (3.1). Here (3.2) applies with L\sim ts, and by (?) one of its two terms is \gtrsim jts. If the term mL/p\sim mts/p is \gtrsim jts, then j\lesssim m/p, so using t\lesssim p and sj\lesssim m,

If the term (pmL)^{1/2}\sim(pmts)^{1/2} is \gtrsim jts, then ts\lesssim pm/j^2, hence s\lesssim m^2/(pj^2) by t>c_0p^2/m, and again ts^2j^4\lesssim p\big(m^2/(pj^2)\big)^2j^4=m^4/p.

The estimate (?) or (?) bounds (?) uniformly for every class. Summing over the classes in (?), combining with (?) and (Proof), and absorbing the resulting O((1+\log m)^4) factor into p^\varepsilon gives (3.3).

Set k_*=m^{2/3}p^{-1/3} and \mu=\min(m,p).

If A and B are k-regular with k\le k_*, then (2.5) gives \mathcal R(A),\mathcal R(B)\lesssim m^2k, so the second term of (?) already gives the result,

Suppose instead that A and B are k-regular with k\ge k_*, and use the first term of (?), \mathcal B(A,B)\le\mathcal T(A,B)=\sum_\theta E_\theta(A)E_\theta(B). Call a pair of distinct points a frame pair if the two points lie on a common frame line and a non-frame pair otherwise, and let F_\theta(A) and E_{\theta,\mathcal L_A}(A) count, respectively, the frame and non-frame pairs of A of direction \theta, where \mathcal L_A is the family of non-frame lines of A. Then E_\theta(A)=F_\theta(A)+E_{\theta,\mathcal L_A}(A), and likewise for B, so expanding the product splits the trapezoid count into three parts,

the frame, non-frame, and mixed contributions, which we bound in turn. The non-frame lines \mathcal L_A and \mathcal L_B meet their sets in O(k)=O(m^{1/2}) points.

We record carefully the number of frame pairs. Write the regular decomposition of A as

where the frame lines \ell_i are distinct. For a fixed i, each set A_j with j\ne i contributes at most one point to \ell_i, because two distinct lines meet in at most one point. Hence

where the last inequality uses k\le m^{1/2}. A distinct pair lies on at most one frame line, and therefore

In particular,

and likewise for B.

For the frame contribution, Cauchy–Schwarz in \theta and (?) give

For the non-frame contribution, Lemma 3.3 applied to (A,\mathcal L_A) and to (B,\mathcal L_B) and Cauchy–Schwarz in \theta give

For the mixed contribution, combining (?) with the bound of Lemma 3.3 for (B,\mathcal L_B) in the form

through Cauchy–Schwarz gives

Using k\ge k_*=m^{2/3}p^{-1/3}, the three summands in (?) are at most

respectively, and each is dominated by the right side of (3.4). For the first term, if m\le p it is m^3p^{1/3}\le m^{8/3}p^{2/3}, while if m\ge p it is at most m^3p^{1/3}=m^3\mu^{1/3}. For the second term, if m\le p it is at most m^{10/3}=m^3\mu^{1/3}; if p\le m\le p^{3/2} it is at most m^3p^{1/3}=m^3\mu^{1/3}; and if m\ge p^{3/2} it is at most m^4/p. For the third term, m^{7/3}p^{5/6}\le m^{8/3}p^{2/3}+pm^2 according as m\ge p^{1/2} or m\le p^{1/2}. The other mixed term, with the roles of A and B exchanged, is identical. Combining the three contributions gives (3.4) for regular sets.

Finally, suppose that A and B are irregular. Let \mathcal L be the family of all lines in \mathbb F_p^2. Since every line meets A and B in at most m^{1/2} points, Lemma 3.3 applies to both (A,\mathcal L) and (B,\mathcal L). Also, because \mathcal L is the family of all lines, every nonzero pair of points of A or B is counted in exactly one line of \mathcal L. Thus

By Cauchy–Schwarz in \theta and Lemma 3.3,

where we used |A|\sim|B|\sim m in the last line. Since \mathcal B(A,B)\le\mathcal T(A,B), this proves the irregular case.

We first dispose of degenerate rectangles. If one of the four edge differences is zero, say x_0=x_1, then the two remaining corner conditions involving the opposite edge give (x_3-x_2)\cdot(x_3-x_2)=0. Since -1 is not a square, the form u\cdot u has no nonzero zeros, and hence x_2=x_3. Thus degenerate rectangles are determined by two points of A, up to cyclic relabeling, and contribute O(n^2), which is absorbed by the right side of (3.5). We may therefore count only rectangles whose edge-lines are genuine lines.

For a dyadic integer k\ge2, let \mathcal L_k denote the lines containing between k and 2k points of A. The Stevens–de Zeeuw incidence theorem gives

On the other hand, Vinh’s estimate gives, whenever k\gtrsim n/p,

We use the standard dyadic rectangle count from the proof of Lemma 2.4. Assign to each rectangle one of its edge-lines having maximal intersection with A. If K\ge1, the rectangles assigned to lines containing at most K points contribute

For a dyadic class k\ge K, the contribution is at most

Indeed, the first line follows by choosing an assigned edge, one endpoint, and the adjacent perpendicular edge. For the second line, the perpendicular lines through the points of a fixed edge are parallel and disjoint, so their intersections with A have total size at most n.

Take a dyadic K\sim C_0p^{3/7}, where C_0 is a sufficiently large absolute constant. Since n\lesssim p^{10/7}, we may choose C_0 so that K\gtrsim n/p. If K\ge n, then (?) already proves the result, so assume K<n. First suppose that n\ge p^{8/7}. Applying (?) in (?), both richness ranges contribute at most

per dyadic class. Together with (?), and after summing the O(\log p) dyadic classes, this proves the claim in this case.

It remains to consider n\le p^{8/7}. For K\le k\le n^{1/2}, combine (?) and (?). We use the elementary inequality

to separate the leading term in (?) from its two error terms. The leading terms satisfy

The other two terms from (?), after multiplication by k^3, are bounded by

where the last inequality uses n\le p^{8/7}. For k\ge n^{1/2}, (?) and the second line of (?) give

where we used k\le n and n\le p^{8/7}. Summing the O(\log p) dyadic classes and absorbing the logarithm into p^\varepsilon proves (3.5).

Proposition 3.4, together with \min(m,p)^{1/3}\le p^{1/3}, gives the right side of (3.6) with m^3p^{1/3} in place of the minimum. Lemma 3.5 and the second alternative in (?) give independently

For nonnegative U,V,D one has \min\{U+V,D\}\le U+\min\{V,D\}. Applying this with

proves the result.

Since (d\sigma)^\vee=\delta_0+K, the contribution of \delta_0 to \mathcal Q(f,h)=\langle f,h*(d\sigma)^\vee\rangle is \langle f,h\rangle, of modulus at most |A\cap B|. For the contribution of K, Hölder’s inequality with exponents \tfrac43 and 4 gives

To estimate \|h*K\|_4 we use the slice estimate of Lemma 2.1 with d=6. For z\in F let h_z(\eta)=h(\eta,z) for \eta\in F^5, supported on B_z=\{\eta\in F^5:(\eta,z)\in B\}, and let \widetilde h_z be its lift to the paraboloid, defined by \widetilde h_z(\eta,\eta\cdot\eta)=h_z(\eta) on its support

so that |\widetilde B_z|=|B_z| and \sum_z|B_z|=|B|. Lemma 2.1 gives

For any function \phi:P\to\mathbb C with |\phi|\le1_E for some set E\subset P, expanding the fourth power and using Plancherel on F^6 identifies \|(\phi d\sigma)^\vee\|_4^4 with a normalized weighted count of additive quadruples in E, the quadruple (\xi_1,\xi_2,\xi_3,\xi_4) with \xi_1+\xi_2=\xi_3+\xi_4 carrying the weight \phi(\xi_1)\phi(\xi_2)\overline{\phi(\xi_3)\phi(\xi_4)}. Since |\phi|\le1_E, each weight has modulus at most 1, so this is at most the unweighted additive energy \Lambda(E), whence

Applying the energy estimate (?) to E=\widetilde B_z and \phi=\widetilde h_z gives \Lambda(\widetilde B_z)^{1/4}\lesssim q^{-1/4}|B_z|^{3/4}+q^{1/2}|B_z|^{1/2}, so combining (Proof) and (Proof),

There are at most q nonempty slices, so Hölder’s inequality gives \sum_z|B_z|^{3/4}\le q^{1/4}|B|^{3/4} and \sum_z|B_z|^{1/2}\le q^{1/2}|B|^{1/2}, and the right side of (Proof) reduces to

Substituting (Proof) into (Proof) gives (4.1), and interchanging the roles of f and h and keeping the smaller of the two bounds gives (4.1). Finally, by Plancherel on F^6,

so the Cauchy–Schwarz inequality |\mathcal Q(f,h)|\le\|\widehat f\|_{L^2(P,d\sigma)}\|\widehat h\|_{L^2(P,d\sigma)} gives (4.1).