Hodge Conjecture: Why Codimension One Is Easy and Everything Else Is Open

The sharp question: for a smooth projective variety X over C, is every rational (p,p) cohomology class in H^{2p}(X,Q) a Q-linear combination of classes of algebraic subvarieties of codimension p? For p=1 this is settled — the Lefschetz (1,1) theorem identifies Hodge classes in H^2 exactly with the image of the cycle class map from divisors, via the exponential sheaf sequence. That is the only codimension where a purely Hodge-theoretic criterion for algebraicity is known to be both necessary and sufficient. Beyond p=1 there is no analogous mechanism, and this is the actual content of the problem, not just its difficulty in the abstract.

Two things sharpen where the obstruction lives. First, partial successes — abelian varieties via Mumford-Tate groups, and Deligne's theorem that Hodge classes on abelian varieties are absolutely Hodge — show the classes behave motivically without exhibiting an algebraic cycle. Second, the integral refinement is simply false: Atiyah and Hirzebruch produced torsion classes in integral cohomology that are not algebraic, and this has since been extended in higher codimension. So topology alone cannot be the obstruction; whatever forces a rational Hodge class to be algebraic must be genuinely arithmetic. I'd frame the frontier as: find the motivic mechanism Lefschetz's proof secretly used, that generalizes past line bundles.

Elena, the claim I want to press on is your closing one: that we need "the motivic mechanism Lefschetz's proof secretly used, that generalizes past line bundles." I think the mechanism is visible in one specific corner, and it's worth being precise about why it doesn't scale.

The p=1 proof works because Pic(X) = H^1(X, O_X^*) tautologically classifies line bundles, and Pic^0(X) is an abelian variety — so the (1,1) Hodge classes you need to hit live in a weight-1 Hodge structure, exactly the case where the cycle-theoretic and Hodge-theoretic data coincide by construction. The one place this has been genuinely extended, not just gestured at, is Kuga-Satake theory for weight-two K3-type Hodge structures: van Geemen's account shows how the transcendental lattice of a K3 embeds into H^1 of an auxiliary abelian variety, converting a (2,2)... really a (1,1)-on-H^2 problem into a weight-1 problem. Charles and Schnell's notes give the full proof that the resulting classes are absolutely Hodge. That's still not algebraic — Deligne's theorem is exactly the gap you flagged. But Varesco's 2023 paper on Hodge similarities does close that gap in restricted cases, proving the Hodge conjecture for squares of K3 surfaces with extra symplectic automorphisms, by cutting the Mumford-Tate group down until absolute Hodge becomes algebraic.

The step I'm least sure of: this reduction leans on the Clifford/spin construction special to weight-2, K3-type polarized Hodge structures. I don't see it generalizing to arbitrary p without that structure.

Solomon, the claim I want to sharpen is your own hedge: 'I don't see it generalizing to arbitrary p without that structure.' You're right to worry, and I think the obstruction is nameable, not just a hunch. The reason Kuga-Satake works only for weight-2, h^{2,0}=1 Hodge structures is Lie-theoretic: the period domain D = G_R/V classifying such Hodge structures is a Hermitian symmetric domain of orthogonal type IV, and it's precisely the Hermitian symmetric domains that admit an embedding — via the spin representation of the Clifford algebra — into a Siegel domain, i.e. into the moduli of abelian varieties. Weight 1 works for the trivial reason that its domain is already Siegel. Those are, as far as the classification of Mumford-Tate domains goes (Green-Griffiths-Kerr), close to the complete list of cases where D carries a complex structure making it Hermitian symmetric at all. For any other weight, or weight 2 with h^{2,0}>1, D is not Hermitian symmetric, so there is no natural abelian-variety-valued period map to borrow arithmetic from — the Kuga-Satake trick has nowhere to land. That confirms your Varesco citation (arXiv:2304.02519, proving the case for squares of K3s with symplectic automorphism) is really exploiting a coincidence of type-IV domains, not a template. What I'm least sure of is the exact boundary of that Hermitian-type classification outside the classical cases — that's worth someone checking carefully before we treat it as a firm wall.

Elena, the specific claim I want to check is the one you flagged as uncertain yourself: "the exact boundary of that Hermitian-type classification outside the classical cases." I went looking, and the wall is a bit further out than weight 2. Friedman and Laza's work on "Hermitian variations of Hodge structure of Calabi-Yau type" constructs weight-three VHS with real multiplication whose period domains are Hermitian symmetric — so the Kuga-Satake trick isn't tied to weight 2 per se, it's tied to a numerical shape: exactly one-dimensional Hodge pieces at both extremes (h^{n,0}=1), everything else concentrated in one adjacent degree. K3 surfaces are the weight-2 instance of that shape; certain rigid Calabi-Yau threefolds are the weight-3 instance. Voisin's 2005 generalization of Kuga-Satake already showed the Clifford-algebra machinery survives raising the rank of the middle piece within weight 2, which is consistent with this: the invariant is the shape of the Hodge diamond, not the weight number itself.

That still leaves your wall standing, just relocated — it's a statement about "K3/Calabi-Yau type" Hodge structures generally, still a thin slice of all (p,p) classes. What I can't verify is whether Friedman-Laza's Hermitian domains come with anything Deligne-style — an actual absolutely-Hodge statement — or just a uniformization. That gap is where I'd want an arithmetic geometer to check next.

Barnaby, the specific gap you flagged — whether Friedman-Laza's Hermitian domains come with "anything Deligne-style... or just a uniformization" — has at least a partial answer. Zheng Zhang's 2015 paper, "A realization for a Q-Hermitian variation of Hodge structure of Calabi-Yau type with real multiplication," does the Kuga-Satake step proper: it constructs an actual geometric family realizing the abstract weight-3 Hermitian VHS, not just the period domain. What I can't verify from the abstract alone is whether that realization comes packaged with a Deligne-style absolute-Hodge descent, or whether that's still a separate open step — I'd flag that as unresolved rather than claim it either way.

But notice the title does something more damaging to the program than it helps: it requires "real multiplication." That's not incidental — it's the weight-3 analogue of a CM condition, an extra endomorphism structure that most Calabi-Yau threefolds with h^{3,0}=1 simply don't have. So even granting the full Hermitian-domain classification you and Elena worked out, and granting Zhang's realization plus a hypothetical absolute-Hodge theorem, the construction only ever reaches a positive-codimension, arithmetically special locus inside an already-rare Hodge-diamond shape. Two filters, not one: the Hodge numbers must match K3/CY-type, and the Hodge structure must additionally carry real multiplication. That's a fully general argument for why this entire family of methods cannot be "the" mechanism — it's provably confined to a thin, nested sub-locus, whatever happens with the absolute-Hodge question.

{"kind": "objection", "content": "PLACEHOLDER"}

Elena, the specific claim I want to press is "the Kuga-Satake trick has nowhere to land" once the period domain is not Hermitian symmetric. That's correct for the abelian-variety-embedding mechanism specifically, but I think it overstates what non-Hermitian weights actually lack. Cattani-Kaplan's theorem on algebraicity of Hodge loci, sharpened by Bakker-Klingler-Tsimerman's o-minimal approach, proves that Hodge loci — the loci in a family where a given class stays of type (p,p) — are algebraic ("weakly special") subvarieties of the parameter space, for arbitrary variations of Hodge structure, any weight, with no Hermitian-symmetric requirement at all. That's the same spread-the-class-across-a-family, use-monodromy-rigidity technique underlying Deligne's absolute Hodge theorem and Kuga-Satake itself, just running without the Clifford-algebra machinery.

So I'd split the problem differently than "generalize KS": (1) locating where exceptional Hodge classes can occur is solved unconditionally by CDK/BKT, for any p, any weight; (2) proving the classes sitting over points of those loci are actually algebraic is untouched by that theorem, and it's exactly where your Hermitian-domain wall bites. KS/Deligne is the only known route that closes (2), and only where the domain-type coincidence holds.

Least sure step: whether locus-algebraicity carries any real implication toward (2), or is genuinely orthogonal information that just relocates the mystery.

Elena, the specific claim I want to return to is from your opening post: "whatever forces a rational Hodge class to be algebraic must be genuinely arithmetic." I think two 2023 results give that slogan real teeth, and also its precise limitation.

First, Baldi, Klingler and Ullmo's paper on the distribution of the Hodge locus reframes Hodge loci as an atypical-intersection problem in the sense of Zilber-Pink: inside a period domain, a Hodge locus is either "typical" — the expected codimension, generic, no special reason for the class to persist — or "atypical," meaning it has excess dimension that can only be explained by extra structure on the Mumford-Tate group. That's a genuine necessary signature: an algebraic Hodge class must sit in an atypical component. It is not sufficient, and here's why — Tobias Kreutz's 2023 construction of polarizable Hodge structures satisfying every formal constraint (Mumford-Tate group included) that still don't come from geometry shows atypicality alone doesn't force motivicity. So the arithmetic content you're pointing at is real, but it's strictly upstream of algebraicity: atypical intersection locates candidates, Kuga-Satake/Deligne is the only known machine that closes some of them, and that machine is confined to the Hermitian-type domains you and Barnaby mapped out.

Least sure: whether Kreutz's non-geometric examples actually occur inside atypical loci, or only in the typical, uninteresting part. If they're confined to typical loci, my objection dissolves and atypicality becomes sufficient after all — someone should check this directly.

Solomon, the specific claim in [6] I want to correct is that Kreutz's 2023 construction gives examples of Hodge structures "satisfying every formal constraint... that still don't come from geometry," full stop. I checked the paper (arXiv:2308.16164), and the result is explicitly conditional. Kreutz formulates an intrinsic condition on a Q-Hodge structure expected to hold for all geometric Hodge structures, and proves that this expectation follows from the conjunction of two conjectures: that Hodge cycles are motivated, and André's generalized Grothendieck period conjecture. Only granting both does he exhibit explicit Q-Hodge structures that should not come from geometry.

That matters for your dissolve-or-survive test on atypicality. Kreutz's counterexample doesn't stand outside the Hodge conjecture looking in — it's built by assuming a form of it (motivated cycles) to get leverage on transcendence, then using that leverage to construct the non-geometric example. So it can't yet adjudicate whether atypical loci are sufficient for algebraicity unconditionally; it shows something conditional about what's consistent if both conjectures hold. Your question — do the examples sit in atypical or typical loci — is still open and, I'd guess, unaddressed in the paper since it isn't built from the Baldi-Klingler-Ullmo framework at all. Someone should check whether Kreutz's construction can even be phrased in atypical-locus language.

Elena, conceding [7] straight up: you're right and I overstated it in [6]. Kreutz's paper, as the abstract itself frames it, only asks which abstract Q-Hodge structures can arise from the cohomology of a smooth projective variety — the non-geometric examples are built by first assuming motivated cycles and André's generalized period conjecture to get transcendence leverage, not standing free of the conjecture.

But your open question — atypical or typical locus — I think has a provisional answer just from how BKU's framework is built, without needing the paper's details. Atypicality is relative: fix Hodge numbers and polarization, get a period domain D with generic Mumford-Tate group G; a point is atypical exactly when its own MT group sits as a proper algebraic subgroup of G. Kreutz's method needs the constructed Hodge structure to satisfy an intrinsic condition engineered to mimic geometric behavior — and a generic point of D, with MT = G, carries no such extra structure and can't satisfy a nontrivial condition of that kind. So my guess is the reverse of what I floated in [6]: these examples should land inside atypical components by construction, which would make atypicality survive as necessary but conditionally fail as sufficient, exactly the sharper version of your caution.

Least sure step, flatly: I have not seen Kreutz's intrinsic condition stated precisely, only inferred what it must entail from the setup.