Deligne (Remark (iii), p. 2): the Hodge conjecture for H2 was proved by Kodaira and Spencer: "a class c∈H2(X,Z) of type (1,1) has image 0 in the quotient H0,2=H2(X,O) of H2(X,C), and the long exact sequence of cohomology defined by the exponential exact sequence 0→Z→O→O∗→0 shows that c is the first Chern class of a line bundle", whose divisor classes are algebraic cycles of codimension one.
Let X⊂PN(C) be non-singular of dimension n≥1 and let c∈H2(X,C) be a Hodge class: rational and represented by a closed (1,1)-form. Then there exist finitely many irreducible subvarieties Z1,…,Zm⊂X of dimension n−1 (divisors), rational numbers ai, and their classes cl(Zi) with
c=i=1∑maicl(Zi).
This is the case p=1 of the goal theorem and the only case known in every dimension; combined with the hard Lefschetz theorem it also gives the case p=n−1, hence the conjecture for dimX≤3.
Formalization Note The statement is the goal theorem specialised to p=1 (degree written 2⋅1, codimension 1), with the same definitions of Hodge class and of cl(Z).
namespace HodgeConjecture
theorem hodge_conjecture_divisors (N n : ℕ) (hn : 1 ≤ n) (X : ProjVariety N)
(hX : X.IsNonsingular n) (c : X.Cohomology (2 * 1)) (hc : X.IsHodgeClass 1 c) :
∃ (m : ℕ) (Z : Fin m → ProjVariety N) (a : Fin m → ℚ) (cl : Fin m → X.Cohomology (2 * 1)),
(∀ i, (Z i).cone ⊆ X.cone ∧ (Z i).IsIrreducible ∧ (Z i).dim = n - 1 ∧
X.IsCycleClass (Z i) n 1 (cl i)) ∧
c = ∑ i, ((a i : ℚ) : ℂ) • cl i := by sorry
end HodgeConjecture
Source
P. Deligne, The Hodge Conjecture, Clay Mathematics Institute Millennium Prize Problem description (2000), https://www.claymath.org/wp-content/uploads/2022/06/hodge.pdf, p. 2, Remark (iii): 'This is the starting point of the proof given by Kodaira and Spencer [7] of the Hodge conjecture for H²'; [7] K. Kodaira, D. C. Spencer, Divisor classes on algebraic varieties, Proc. Nat. Acad. Sci. 39 (1953), 872–877
Read-back
What the Lean code literally says, in plain math · claude-fable-5-1
Read-back: hodge_conjecture_divisors
Ambient objects. Fix a natural number N. Write V:=CN+1 (functions {0,…,N}→C) and A:= the space of (N+1)×(N+1) complex matrices, viewed as the complex Hilbert space ℓ2 on index pairs (i,j) (Frobenius inner product ⟨A,B⟩=∑ijAijBij); as a real vector space A has dimension 2(N+1)2 and carries the induced Euclidean metric. Let P:=C[x0,…,xN].
The following auxiliary maps are used throughout.
proj:V→A, proj(v)ij=vivj/∑k∣vk∣2 (with the convention that division by 0 gives 0, so proj(0)=0). For v=0 this is the rank-one orthogonal projection onto Cv.
JP(A):=i(AP−PA)=i[A,P] for P,A∈A.
cmulP(z,A):=(Rez)A+(Imz)JP(A) for z∈C.
⟨a,b⟩R:=Re⟨a,b⟩ on A.
For a set S⊆P, zeros(S):={v∈V:Q(v)=0∀Q∈S}.
Projective varieties. A ProjVarietyX (of ambient index N) is a set of polynomials X.eqs⊆P together with a proof that each Q∈X.eqs is homogeneous of some degree dQ∈N (degrees may differ; the zero polynomial qualifies; the set may be empty, or may contain a nonzero constant, in which case the cone below is empty). Two varieties with different equation sets but the same zero set are different objects. For such X:
cone(X):={v∈V:v=0,Q(v)=0∀Q∈X.eqs} (nonzero vectors of the affine cone).
pts(X):=proj(cone(X))⊆A.
van(X):={Q∈P:Q(v)=0∀v∈cone(X)}, all polynomials (homogeneous or not) vanishing on the cone.
For Q∈P and v∈V, dQv:V→C is the C-linear map u↦∑i∂iQ(v)ui.
TvX:=⋂Q∈van(X)ker(dQv)⊆V, a complex subspace (defined for every v∈V, not only v∈cone(X); it is cut out by the full vanishing ideal, not just by X.eqs).
IsNonsingular(X,n): for every v∈cone(X), dimCTvX=n+1. (Vacuous if the cone is empty.)
tanX(P)⊆A, a real subspace: the span (supremum of subspaces) over all v∈cone(X) with proj(v)=P of the image of TvX (viewed as a real subspace) under the real Fréchet derivative Dproj(v):V→A. If P∈/pts(X) the supremum is over an empty family and tanX(P)={0}. (The Fréchet derivative is a total function equal to 0 where proj is not differentiable; proj is differentiable at every v=0.)
IsIrreducible(X): cone(X)=∅, and for all sets S1,S2⊆P (arbitrary, not necessarily homogeneous), if cone(X)⊆zeros(S1)∪zeros(S2) then cone(X)⊆zeros(S1) or cone(X)⊆zeros(S2).
dimX:=(infv∈cone(X)dimCTvX)−˙1, where −˙ is truncated natural-number subtraction and the infimum of an empty set of naturals is 0. So an empty cone gives dimX=0, and dimX=0 also whenever the infimum is 0 or 1.
reg(X):=proj({v∈cone(X):dimCTvX=dimX+1}), the images of cone points where the tangent-space dimension attains the minimum.
Forms and cohomology. A k-form is any function w:A→AltRk(A;C) assigning to each point of the whole ambient space A a continuous alternating R-multilinear map Ak→C. Forms are ambient objects; X enters only via the following restrictions.
wvanishes onX: for every P∈pts(X) and every k-tuple v1,…,vk∈tanX(P), wP(v1,…,vk)=0.
w is closed onX: w is C∞ on all of A and its exterior derivative dw (Mathlib's total exterior derivative of a form, a (k+1)-form) vanishes on X.
w is exact onX: for k=0, w is C∞ and vanishes on X; for k≥1, w is C∞ and there exists a C∞(k−1)-form α such that w−dα vanishes on X.
Zk(X):= the C-linear span of the closed k-forms; Bk(X):= the C-linear span of the exact k-forms.
Hk(X):= the image of Zk(X) in the quotient {all k-forms}/Bk(X), a complex vector space; [w]∈Hk(X) denotes the class of w∈Zk(X).
whas type (p,q) onX: for every P∈pts(X), every tuple v1,…,vk∈tanX(P), and every z∈C,
A class c∈Hk(X)has type (p,q) if there exists some w∈Zk(X) with [w]=c and w of type (p,q) on X.
Chains, periods, rationality. A k-simplex is any function σ:Rk+1→A (no regularity assumed at this stage); a k-chain is a finite formal Z-combination ∑σmσσ of simplices. The i-th face of a (k+1)-simplex is t↦σ(t with a 0 inserted in slot i); ∂(∑mσσ):=∑σmσ∑i=0k+1(−1)ifaceiσ. A chain is a cycle if k=0 (always) or ∂c=0 (for k≥1). With Δk:={t∈Rk:ti≥0,∑ti≤1} and e(t):=(1−∑iti,t1,…,tk)∈Rk+1,
a Bochner integral against Lebesgue measure on Rk, where εi are the standard basis vectors, D is the (total) real Fréchet derivative, taken to be 0 where σ∘e is not differentiable, and the integral is 0 if the integrand is not integrable. For k=0, Δ0 is a single point of mass one and ∫σw=wσ(1). The period is per(w,c):=∑σmσ∫σw∈C.
A simplex σ is inX if σ is C∞ on all of Rk+1 and σ(t)∈pts(X) for every t in the standard simplex {ti≥0,∑ti=1}; a chain is in X if every simplex in its support is.
IsRational(X,c) for c∈Hk(X): for everyw∈Zk(X) with [w]=c and everyk-chain ch that is in X and is a cycle, there exists q∈Q with per(w,ch)=q.
IsHodgeClass(X,p,c) for c∈H2p(X): IsRational(X,c)andc has type (p,p).
Frames, measures, pairing. For a real subspace T⊆A, a point P∈A and u:{0,…,m−1}→A, u is a J-frame in T at P if all ui∈T, ⟨ui,uj⟩R=δij, and ⟨ui,JPuj⟩R=0 for all i,j. The framefr(T,P,d)∈Ad is built as follows: let u∈A⌊d/2⌋ be someJ-frame of length ⌊d/2⌋ in T at P chosen by Hilbert's ε (Classical.epsilon) — if none exists, u is an arbitrary, unspecified tuple with no properties — and set
fr(T,P,d)k:=⎩⎨⎧u⌊k/2⌋JPu⌊k/2⌋0k even,⌊k/2⌋<⌊d/2⌋,k odd,⌊k/2⌋<⌊d/2⌋,otherwise (only when d is odd and k=d−1).
The measure μd:=⌊d/2⌋!2dπ⌊d/2⌋⋅Hd on A, where Hd is Mathlib's d-dimensional Hausdorff measure for the Euclidean metric of A. For a k-form w, an l-form η, a point P and f∈Ak+l,
both Bochner integrals over the indicated sets (restriction of the measure to a possibly non-measurable set, via outer measure), each equal to 0 if the integrand fails to be integrable. Note the pairing integrates over pts(X) while ∫Z integrates over reg(Z) only.
IsCycleClass(X,Z,n,p,c) for c∈H2p(X): for every w∈Z2p(X) with [w]=c and every (2n−˙2p)-form η that is closed on X,
⟨w,η⟩X=∫Zη.
Here 2n−˙2p is truncated subtraction. The wedge in the pairing is a (2p+(2n−˙2p))-form evaluated on a frame of that length in tanX(P).
The theorem. Let N,n∈N with 1≤n, let X be a ProjVariety of ambient index N satisfying IsNonsingular(X,n) (every cone point has dimCTvX=n+1), let c∈H2(X) — the degree is written 2⋅1 — and assume IsHodgeClass(X,1,c), i.e.
every closed representative w of c has a rational period over every smooth 2-chain in X that is a cycle, and
some closed representative w of c satisfies wP(cmulP(z,v1),cmulP(z,v2))=zzˉwP(v1,v2)=∣z∣2wP(v1,v2) for all P∈pts(X), all v1,v2∈tanX(P), all z∈C.
Then there exist a natural number m, a family of ProjVarieties Z0,…,Zm−1 of ambient index N, rationals a0,…,am−1∈Q, and classes cl0,…,clm−1∈H2(X) such that
for every i<m:
cone(Zi)⊆cone(X);
IsIrreducible(Zi) (in particular cone(Zi)=∅);
dimZi=n−˙1, i.e. (infv∈cone(Zi)dimCTvZi)−˙1=n−1 (genuine subtraction since n≥1);
IsCycleClass(X,Zi,n,1,cli): for every closed 2-form w on X with [w]=cli and every (2n−2)-form η closed on X, ⟨w,η⟩X=∫Ziη, where the left side uses a frame of length 2n in tanX(P) over pts(X) with measure μ2n, and the right side uses a frame of length 2n−2 in tanZi(P) over reg(Zi) with measure μ2n−2;
and, in the complex vector space H2(X),
c=i<m∑ai⋅cli,
with each ai∈Q coerced to C.
No nonsingularity, no relation between cli and any specific form, and no distinctness of the Zi is required; the only link between Zi and cli is the IsCycleClass condition. The parameter n enters IsCycleClass independently of dimX (which is never referenced for X).
Edge and degenerate cases visible in the statement.
Trivial witness. With m=0 the family conditions are vacuous and the sum is 0; so the conclusion holds for c=0 regardless of hypotheses. More generally any cli that is the zero class of H2(X) satisfies IsCycleClass with respect to some Zi exactly when ⟨w,η⟩X=∫Ziη for all exact-on-X representatives w of 0 and closed η.
Empty variety. Nothing forbids cone(X)=∅ (e.g. a nonzero constant among the equations). Then IsNonsingular is vacuous, pts(X)=∅, every form vanishes on X, every C∞ form of positive degree is exact (take α=0), so H2(X)=0, c=0, and the conclusion is satisfied with m=0. Irreducibility of the Zi forces cone(Zi)=∅, so with m≥1 the subvarieties would have to be nonempty.
Hypothesis strength. IsRational quantifies over all representatives and all smooth cycles in X (the empty chain is always such a cycle, with period 0); HasTypeClass asks for one representative of type (1,1). IsCycleClass quantifies over all representatives w of cli and all ambient (2n−2)-forms η that are closed on X.
Junk values. The pairing and ∫Z are Bochner integrals returning 0 when the integrand is not integrable (or not a.e. strongly measurable) with respect to the scaled Hausdorff measure restricted to pts(X) or reg(Z); the exterior derivative and Fréchet derivatives are total functions returning 0 at non-differentiable points; the frame is chosen by Classical.epsilon and is an arbitrary tuple whenever no J-frame of length ⌊d/2⌋ exists in the given tangent space (e.g. when tan(P) has too small a real dimension or is not JP-stable), in which case the integrands are evaluated on unspecified vectors; for odd frame length the last vector is 0 (here the lengths 2n and 2n−2 are even).
Natural subtraction.n−˙1 and 2n−˙2 are genuine because 1≤n; for n=1 the test forms η are 0-forms (functions A→C), the frame on Zi has length 0, and ∫Ziη=∫reg(Zi)ηdμ0 with μ0=H0 (counting measure). In dimZi, if the infimum of tangent dimensions were 0 or 1 the truncated result is 0, which meets the requirement dimZi=n−1 exactly when n=1; the infimum over an empty cone is excluded by irreducibility.
Scalars. Cohomology is a C-vector space; the coefficients ai are rationals cast into C, so the conclusion asserts membership of c in the Q-span of the classes cli.
Tangent-space conventions.tanX(P) is a real subspace of the ambient matrix space obtained by pushing the Zariski tangent space of the affine cone (computed from the full vanishing ideal) through Dproj; "closed", "exact", "type (p,q)", and "vanishes" are all tested only against these subspaces at points of pts(X), while smoothness of forms and simplices is demanded on the entire ambient space A, resp. all of Rk+1.