Computing Picard Lattices of K3 Surfaces

Lecture 2: p-adic Hodge-theoretic obstructions

Edgar Costa (MIT)

ICERM: Arithmetic, Geometry and Computations on K3 surfaces
September 16, 2026

Supported by the Simons Foundation

Slides available at edgarcosta.org

Where we got to yesterday

Reduction gives

ρ(Xal) ≤ρ(Xp) for every good p.

  • Jumping is frequent: a non-trivial jump character makes a density-one-half set of primes overshoot.
  • Forced excess is different: every good prime overshoots. Exactly two cases:
    • EX = ℚ, dim TX odd: forced excess η= 1, van Luijk survives;
    • EX totally real, EX ≠ℚ, dim EX TX odd: excess [EX:ℚ] ≥2, a wall.

Only the second defeats the unaugmented two-prime argument.

  • T=T(X); E=End Hdg(T); d=[E:ℚ]; m=dim E T
Theorem (Charles)

ρ(Xpal)≥{ρ(Xal)if E is CM or m is even,ρ(Xal)+dif E is totally real and m is odd,

Equality occurs infinitely often (density 1 after some finite extension).

If E is totally real and m is odd, infinitely many good ordinary prime pairs (p,q) satisfy ρ(Xpal)=ρ(Xqal)=ρ(Xal)+d and

disc Pic (Xpal)≢disc Pic (Xqal) mod (ℚ×)2

It is very hard to prove RM!

We are now not only asking to find algebraic cycles in X but also in X ×X.

Picard lattice, over finite fields

Xp/𝔽q a K3 surface; q=pn, ℓ∤q.

P2(t) := det (t − Frob q ∣H2et(Xpal, ℚ)) ∈ℤ[t].

Tate conjecture

Pic (Xp) = ker ( Frob q − q ·id ∣H2et(Xpal, ℚ))

Tate conjecture is known for K3 surfaces over finite fields.

Since Frob q acts semisimply, we have:

ρ(X𝔽qm) = ∑ζm=1 ord t=qζ P2(t).

ρ(Xpal) = ∑ζ ord t=qζ P2(t),

where ζ runs over all roots of unity.

Note: ρ(Xpal) ≡0 mod 2

Picard lattice, over finite fields

det (1 − tFrob q ∣H2et(Xpal, ℚ)) = t22P2(1/t).

The Hasse-Weil zeta function ZXp(t) can be written as

ZXp(t) := exp ( ∑m=1 (# Xp(𝔽qm))/(m) tm ) = (1)/((1−t) t22P2(1/t) (1−q2 t)).

One may deduce P2 from the point counts # Xp(𝔽qm) for m ≤b2/2 + 1 = 12.

Reduction to finite characteristic

In this lecture we will focus on projective hypersurfaces, with quartic K3s in mind, but the methods are more generic, as at some point we will need to do explicit computations.

Take f ∈ℤ[x,y,z,w] and X := Z(f) ⊂ℙ3.

We may consider the surface X𝔽p := Z(f mod p) ⊂ℙ3(𝔽p).

Yesterday, we mostly counted classes, but we also saw the cokernel theorem.

Theorem

If X and X𝔽p are smooth then the specialization map is injective

Pic (Xal) ↪Pic (X𝔽pal)

and ρ(Xal) = ρ(Xpal) ≤ρ(X𝔽pal).

The specialization map has torsion-free cokernel for p ≠2.

Can we use it without computing Pic (X𝔽pal)?

The lifting question

Goal

For a given f and p, improve the inequality ρ(Xal) ≤ρ(Xpal).

Which classes in Pic (Xpal) actually come from X?

  • Every excess counted yesterday is a class that exists in the special fibre and doesn't lift.
  • Reduction ranks cannot see the difference: they count the special fibre and nothing else.

We will do this by considering the thickenings

Z(f mod pi) ⊂ℙ3ℤ/(p)i i = 1, 2, ...

Not "which prime" but "which classes": lifting, not counting.

1st ingredient: cohomology

Choose a finite extension K'/ℚp, with residue field k'=𝔽pm, so the geometric divisor classes of X are defined over K' and those of Xp over k'.

Over characteristic zero we have:

  • H:=H2dR(X/ℚp) = F0 ⊃F1 ⊃F2, the Hodge filtration
  • Pic (Xal) ↪F1K':=F1pK'
  • For d = 4, dim Fi(X) = 22, 21, 1.

Over characteristic p we have:

  • Pic (Xpal) ↪H2crys(Xp×𝔽pk'/W(k'))⊗W(k')K' ≃H⊗pK'=:HK'=F0K'⊃F1K'⊃F2K'

Berthelot-Ogus-Raynaud

Theorem (Berthelot-Ogus-Raynaud)
  • Pic (Xpal)⊂HK'
  • sp (Pic (Xal))=Pic (Xpal)∩F1K'

Abelian surface

A = Jac (y2 = 4x5 − 36x4 + 56x3 − 76x2 + 44x − 23)

L(t) = det (1 − tFrob ∣H1) = 1 − 3t + 14t2 − 93t3 + 961t4.

Frob |H1dR(A/ℚp) ≡(31 ·48231 ·28416241307531 ·38631 ·88626441212631 ·28431 ·6596336975031 ·19431 ·8762740810841) (mod 313),

From this we deduce Frob |H2dR(A/ℚp) and

det (1 − t 31−1Frob ∣H2dR(A/ℚp)) = (t−1)2(31t4 + 48t3 + 43t2 + 48t + 31)/31

Thus, ρ(A𝔽pal) = 2.

Since the basis of H1 respects the Hodge filtration, the induced basis in H2 will also respect it.

Abelian surface, continued

SageMath package: crystalline_obstruction.

A = Jac (y2 = 4x5 − 36x4 + 56x3 − 76x2 + 44x − 23)

det (1 − t 31−1Frob ∣H2dR(A/ℚp)) = (t−1)2(31t4 + 48t3 + 43t2 + 48t + 31)/31

Thus, ρ(A𝔽pal) = 2.

Compute 2 eigenvectors

v1(356, 37, 831, 0, 295, 31) (mod 312)v2(4, 957, 3, 1, 0, 0) (mod 312).

The last coordinate of the vectors above gives the projection to H2/F1. Therefore, v1 ∉F1 and the corresponding algebraic cycle cannot lift to p.

Thus, we improved rank NS (Aal) ≤2 to rank NS (Aal) ≤1, and therefore End (Aal) = ℤ.

van Luijk's method would have succeeded in this example by using a second prime.

What Frobenius acts on

Via the isomorphism H2crys(X𝔽p/ℤp) ⊗ℚp ≃H2dR(X/ℚ), we have

Frob p : H2dR(X/ℚp) →H2dR(X/ℚp).

The obstruction map [C-Sertöz]

Compute a p-adic approximation of the obstruction map

π: Pic (X𝔽p) ⊂H2crys(X/ℤp) ⟶H2crys(X/ℤp) / F1 H2crys(X/ℤp)

If π(C) ≠0, then C ∉Pic (X).   (analogous to Pic (X) = H1,1(X) ∩H2(X, ℤ))

  1. compute a p-adic approximation of Frob p
  2. compute an approximation of Pic (X𝔽p)p = ker ( Frob p − p ·id ∣H2dR(X/ℚp))
  3. compute an approximation of πp : Pic (X𝔽p)p →H2dR(X/ℚp) / F1 H2dR(X/ℚp)
  4. dim Pic (X) ≤dim p ker πp

By picking a basis that respects the Hodge filtration, the map H2dR(X/ℚp) →H2dR(X/ℚp)/F1p is a coordinate projection.

What you actually compute

  • Frobenius is not known exactly.
  • Today we use a p-adic toric approach to compute Frobenius modulo pN, for a chosen precision N.
  • For a hypersurface, use a basis of differential forms on its affine complement.
  • Apply Frobenius and reduce its images back to the basis to obtain the matrix modulo pN.
  • What matters here is the shape of the output: an approximation whose error you control.

Why finite precision still proves something

  • An approximate Frobenius gives an approximate eigenspace.
  • But π(C) ≠0 is an open condition.
  • Establishing it to finite precision establishes it.
  • Every dimension the method removes is removed rigorously: a genuine upper bound at any N.

Raising N can only remove more. It never puts a dimension back.

K3 surface

X := Z(y4 − x3 z + y z3 + z w3 + w4) ⊂ℙ3

p=89, N=3, F=89−1Frob 89

  • +1-eigenspace, polarization: dimension 1. No obstruction; this class lifts.
  • +1-eigenspace, primitive part: dimension 4. Obstruction found; at most 3 dimensions remain.
  • −1-eigenspace: dimension 1. Obstructed; no dimension remains.
  • Primitive eighth-root piece ker (F4+1): dimension 4. Obstructed; no dimension remains.
  • ρ(X𝔽89al)=5+1+4=10, ρ(Xal)≤(1+3)+0+0=4

In fact, ρ(Xal) = 4 as there are four lines in z = 0.

previous approaches would not have used p = 89

at p = 89 the obstruction improves 10 to 4, and 4 is sharp;

Does this always work?

Quartic surface

X = Z(y4 − x3 z + y z3 + z w3 + w4) ⊂ℙ3

p=31, N=5

factor of p−1Frob pdimensionliftable dimension, at most
t−111
t−111
(t+1)222
  • ρ(X𝔽31al) = 4
  • no cycle obstruction found while working ℤ/(p)5

by searching for lines Elsenhans-Jahnel's method would have succeeded in this example

at p = 31 the reduction bound is already the sharp value 4, so there is nothing left to remove;

Quintic surface

X := Z(9 x y4 + 3 x4 z + 9 y2 z3 + z5 + 5 w5) ⊂ℙ3

p=23, N=6, ρ(Xal) ≤1

factor of p−1Frob pdimensionliftable dimension, at most
t−111
t−110
t+110
t2+120

ρ(Xal) = 1

p=29, N=20, ρ(Xal) ≤3

factor of p−1Frob pdimensionliftable dimension, at most
t−111
(t−1)221
(t+1)221

The surface has CM by ℚ(ζ5)

at p = 29 on the quintic the obstruction fires, improving 5 to 3, but 3 is not sharp and no precision brings it to 1.

What the three examples say

Three outcomes, and they are different in kind:

  • at p = 89 the obstruction improves 10 to 4, and 4 is sharp;
  • at p = 31 the reduction bound is already the sharp value 4, so there is nothing left to remove;
  • at p = 29 on the quintic the obstruction fires, improving 5 to 3, but 3 is not sharp and no precision brings it to 1.

Only the third is a statement about the method rather than about an attempt.

So the open question is not whether some prime works, but whether one can prove that some prime works.

Does this always work?

No.

Is there a prime for which the bound will be tight?

What is being computed, exactly

Can we combine both approaches?

  • At the moment we only certify an upper bound for dim pL, where L is the largest Frob p-stable subspace of Tev∩F1.
  • To combine several primes we need at least Pic (Xpal), to be able to use sp (Pic (Xal))=Pic (Xpal)∩F1K' in its full strength.
  • At the moment we are only using sp (Pic (Xal))⊗K'⊆L⊗pK'⊆(TevpK')∩F1K'

Raising N does not touch the missing rational structure.

We would also love to combine this with methods over .

Theoretical example

X : w2 = (−y2/8 + yz − z2)(7x2/8 + 5xz + 7z2)(2x2 + 3xy + y2)

  • The real multiplication example from Lecture 1: X is the minimal resolution of this double cover of 2; the sextic is a product of three conics, singular at 15 points.
  • Known sublattice: the polarization and the 15 exceptional curves, so ρ(Xal)≥16.
  • At p = 83 it has χ1 = (t−1)10(t+1)6, and the reduction bound leaves ρ(Xal)=16,17 or 18.
  • The two extra classes span the piece t2+1, irreducible over 83, so 17 is out: ρ(Xal)=16 or 18. In fact ρ(Xal)=16, and X has RM by ℚ(√(2)).
  • This is the wall from Lecture 1: RM by a field of degree 2 with (22 − 16)/2 = 3 odd, so every good prime overshoots.
  • Over , a nonzero obstruction on the t2+1 piece would remove both extra dimensions and prove ρ(Xal)=16, without knowing RM.
  • After base change to ℚ(√(2)), the residue degree at 83 is 2. Using only F2, where F=83−1Frob 83, gives F2=−1 on the extra piece. A kernel line is Frobenius-stable and survives: the bound is at least 17.

"Given a good enough approximation to Frob p we will obstruct 2 extra cycles. However, we don't yet know how to compute such approximation!"

The question

Is there a prime for which the bound will be tight?

  • Does it give sharp bounds? Yes: the quartic at p=89, from 10 to 4.
  • Does it always improve the reduction bound? No: at p=31 the quartic already has the sharp bound 4.
  • Does more precision always make the bound sharp? No: the quintic at p=29 stays at 3, although its Picard number is 1.
  • Can base change lose sharpness? Yes: the real multiplication example at 83 would give 16 over , but at least 17 using only the squared Frobenius over ℚ(√(2)).

Is there always a sharp prime over ?

Lecture 2