# The Hodge conjecture for products of K3 surfaces with Picard number at least 17, and in Floccari's locus

## Abstract

**Theorem 1.** Let $S_1,\dots,S_r$ be projective K3 surfaces with Picard numbers $\rho(S_i)\ge17$. Then the Hodge conjecture holds for every product $S_1^{a_1}\times\cdots\times S_r^{a_r}$.

**Theorem 3.** HC holds for every product of powers of K3 surfaces each of which either has $\rho\ge17$, or has endomorphism field $\mathbf Q$ and transcendental space embedding in $(U^3\oplus\langle-m\rangle)_{\mathbf Q}$ (Floccari's locus). The second class includes every surface of Picard number 16 with $E=\mathbf Q$ and Witt index 2.

The products may mix surfaces of different Picard numbers with non-isomorphic transcendental lattices, and their Hodge rings contain classes not of the form $\mathrm{Hom}_{\rm Hdg}(T(S_i),T(S_j))$; an explicit mixed class is given in Example 6.1.

**Proof outline.**
- Every rational K3-type space of rank $\le5$ embeds in $(U^3\oplus\langle-m\rangle)_{\mathbf Q}$. By Floccari's theorem, the Kuga–Satake correspondence of each $S_i$ is therefore algebraic.
- An algebraic retraction from Hard Lefschetz and Hodge–Riemann turns this into an algebraic direct-summand embedding of $T(S_i)$.
- The Kuga–Satake isogeny factors that occur either have Hodge groups of types $A_1$ and $C_2$ with standard representations, or are CM abelian varieties of dimension $\le2$. A Goursat argument and the first fundamental theorem for $\mathrm{Sp}$ show that the Hodge ring of any product of the first kind is generated by divisors. For the CM factors we use Ramón-Marí's theorem for products of abelian varieties of dimension $\le2$ [1, Prop. 2.18], and the two kinds do not interact.

For Theorem 3 we add two inputs. First, Varesco's theorem makes every Hodge similarity between two such surfaces algebraic. Second, a Goursat argument at the level of $T$ shows that for $\mathrm{rk}\,T=6$ there are no mixed classes other than similarities (§6A).

Among non-CM surfaces with endomorphism field $\mathbf Q$, genuinely exceptional mixed classes come only from split rank-4 factors (Remark 6A.3). The Kuga–Satake method of Proposition 5.2 does not extend beyond rank 5: in rank 6 the analogous Hodge ring has exceptional degree-4 classes. For $\rho\ge19$ the result also follows from Shioda–Inose structures and is presumably known to experts. For $\rho\in\{17,18\}$ we have not found it in the literature.

## 1. Setting

$T(S)=\mathrm{NS}(S)^\perp\subset H^2(S,\mathbf Q)$ denotes the rational transcendental Hodge structure, $\psi$ the cup product, and $E=\mathrm{End}_{\rm Hdg}(T)$. By Zarhin [2], $E$ is a totally real or CM field, and:
- if $E$ is totally real, then $\mathrm{Hg}(T)=\mathrm{Res}_{E/\mathbf Q}SO(T,\phi)$ for an $E$-bilinear form $\phi$, and $\dim_ET\ge3$; in particular $\mathrm{Hg}(T)=SO(T)$ if $E=\mathbf Q$;
- if $E$ is CM, then $\mathrm{Hg}(T)=\mathrm{Res}_{E^+/\mathbf Q}U(T,\phi)$ for an $E/E^+$-Hermitian form $\phi$, and $\dim_ET\ge1$ (see also [3, §2.1]).

A CM endomorphism field does not force a commutative Hodge group: for $\dim_ET\ge2$ the unitary group has a nontrivial semisimple part.

Hence $\rho\ge17$ (i.e. $\mathrm{rk}\,T\le5$) leaves exactly the following cases. Odd rank forces $E$ to be totally real, hence $E=\mathbf Q$; in rank 4 a totally real $E\ne\mathbf Q$ would need $\dim_ET\ge3$, and a CM field has degree 2 or 4.

| $\rho$ | $\mathrm{rk}\,T$ | $E$ | $\mathrm{Hg}(T)$ |
|---|---|---|---|
| 20 | 2 | imaginary quadratic $K$ | $U_K=\ker(N\colon R_{K/\mathbf Q}\mathbf G_m\to\mathbf G_m)$ |
| 19 | 3 | $\mathbf Q$ | $SO(T)$, form of $PGL_2$ |
| 18 | 4 | $\mathbf Q$ | $SO(T)$ |
| 18 | 4 | imaginary quadratic $K$, $\dim_KT=2$ | $U(T,\phi)$, $\phi$ Hermitian of signature $(1,1)$ |
| 18 | 4 | quartic CM, $\dim_ET=1$ | $U_E$, a torus of dimension 2 |
| 17 | 5 | $\mathbf Q$ | $SO(T)$ |

## 2. Rank $\le5$ lies in Floccari's locus

**Lemma 2.1** ([4, Prop. 3.5], proof reproduced). Every rational quadratic space $T$ of rank $r\le5$ and signature $(2,r-2)$ embeds isometrically in $(U^3\oplus\langle-m\rangle)_{\mathbf Q}$ for some integer $m>0$.

*Proof.*
- Replace $T$ by $T\oplus\langle-1\rangle^{5-r}$, so $r=5$ and $\delta=\det T<0$.
- Diagonalise $T=\langle a_1,\dots,a_5\rangle$ over $\mathbf Q$. Let $S$ be $\{2\}$ together with the primes dividing a numerator or denominator of some $a_i$. Pick, by the Chinese remainder theorem and Dirichlet's theorem, a prime $\ell\notin S$ that is a non-square in $\mathbf Q_p$ for all $p\in S$. Let $m$ be the positive squarefree integer in the square class of $-\delta\ell$ (its prime factors lie in $S\cup\{\ell\}$), and put $V=U^3\oplus\langle-m\rangle$.
- A complement $C$ is binary of signature $(1,1)$ with $\det C\equiv e:=-\ell$ modulo squares, and its Hasse invariants $\varepsilon_p$ are forced by $c_p(V)=c_p(T)\,\varepsilon_p\,(\delta,e)_p$. For $C=\langle x,e/x\rangle$ one has $c_p(C)=(x,e/x)_p=(x,-e)_p$. By [5, IV, Prop. 7], $C$ exists iff the following hold:
  - $\prod_p\varepsilon_p=1$. This follows from the product formula.
  - $\varepsilon_p=1$ wherever $-e=\ell$ is a square in $\mathbf Q_p$. That never happens for $p\in S\cup\{\ell\}$. For odd $p\notin S\cup\{\ell\}$, all diagonal entries of $T$ and $V$ are $p$-adic units, so $\varepsilon_p=1$.
  - The real data match: $V_{\mathbf R}=T_{\mathbf R}\oplus\langle1,-1\rangle$.
- Hasse–Minkowski gives $T\oplus C\cong V$. ∎

**Corollary 2.2.** For $\rho(S)\ge17$, the Kuga–Satake correspondence $\mu'\colon H^2(S,\mathbf Q)_{\rm prim}\hookrightarrow H^2(KS(S)^2,\mathbf Q)$ of [6, (2)] is induced by an algebraic cycle, and HC holds for all powers of $S$.

*Proof.* Lemma 2.1 and [6, Thm 5.11]. ∎

`rank5_floccari.py` checks the construction of Lemma 2.1 on 3000 random spaces.

## 3. An algebraic retraction

**Lemma 3.1.** Let $S$ be a projective K3 surface, $A$ an abelian variety of dimension $n$, and $\iota\colon T(S)\hookrightarrow H^2(A,\mathbf Q)$ an injective morphism of Hodge structures induced by an algebraic cycle. Then there is an algebraic correspondence $r\colon H^2(A,\mathbf Q)\to T(S)$ with $r\circ\iota=\mathrm{id}_{T(S)}$.

*Proof.*
- **The pairing.** Let $L$ be an ample class on $A$, and put $B(x,y)=\int_Ax\,y\,L^{n-2}$ on $H^2(A)$. Its restriction to $\iota(T)$ is a Hodge-invariant symmetric form.
- **It is nonzero.** Let $0\ne\sigma\in T^{2,0}$. Then $x=\iota(\sigma)\in H^{2,0}(A)$ is nonzero. It is primitive, because $L^{n-1}x$ has type $(n+1,n-1)$ in degree $2n$ and so vanishes. By Hodge–Riemann, $B(x,\bar x)\ne0$.
- **It is nondegenerate.** Its radical is a sub-Hodge structure of $T$, which is irreducible [2], and is proper because $B(x,\bar x)\ne0$. So the radical is $0$.
- **The adjoint.** Define $\iota^*\colon H^2(A)\to T$ by $\psi(\iota^*y,t)=B(y,\iota t)$. It is induced by the transpose of $\iota$ composed with $L^{n-2}$ and the algebraic projector $\pi_T$ of $S$, so it is algebraic.
- **The retraction.** $\iota^*\iota\in\mathrm{End}_{\rm Hdg}(T)=E$ is nonzero, hence invertible in the field $E$. Its inverse is a polynomial in $\iota^*\iota$ with rational coefficients, so $r=(\iota^*\iota)^{-1}\iota^*$ is algebraic. ∎

Applied to Corollary 2.2, with $A=KS(S)^2$ and $\iota=\mu'|_{T}$, this exhibits $T(S)$ as an algebraic direct summand of $H^2(KS(S)^2)$.

## 4. The Kuga–Satake factors

Write $H^2(S,\mathbf Q)_{\rm prim}=T\oplus N$ with $N$ of Tate type.
- **Hodge group.** By [7, §4] and [6, Rem. 3.2], $\mathrm{Hg}(KS(S))$ is the identity component of the preimage of $\mathrm{Hg}(T)$ in $\mathrm{Spin}(T\oplus N)$; since $N$ is of Tate type this is the identity component of the preimage of $\mathrm{Hg}(T)$ in $\mathrm{Spin}(T)$. For $E=\mathbf Q$ this is $\mathrm{Spin}(T)$.
- **Representation.** The spin module of $\mathrm{Spin}(T\oplus N)$ restricts to $\mathrm{Spin}(T)$ as a sum of (half-)spin representations.

Over $\bar{\mathbf Q}$ this gives:

| $\mathrm{rk}\,T$ | $\mathrm{Hg}(KS)_{\bar{\mathbf Q}}$ | $H^1(KS)_{\bar{\mathbf Q}}$ |
|---|---|---|
| 3 | $SL_2$ | $\mathrm{std}^{\oplus m}$ |
| 4 | $SL_2\times SL_2$ | $(\mathrm{std}\boxtimes1)^{\oplus m}\oplus(1\boxtimes\mathrm{std})^{\oplus m'}$ |
| 5 | $\mathrm{Sp}_4$ | $\mathrm{std}^{\oplus m}$ |
| 2 (CM by $K$) | $U_K$ | $(z\oplus\bar z)^{\oplus m}$, i.e. $KS\sim E_K^{m}$ |
| 4, $E=K$ imag. quadratic, $\dim_KT=2$ | $SL_2\times\mathbf G_m$ | $(\mathrm{std}\boxtimes1)^{\oplus m}\oplus(1\boxtimes(z\oplus z^{-1}))^{\oplus m}$ |
| 4, $E$ quartic CM | 2-dimensional torus | four distinct characters, each with multiplicity $m$ |

*The two CM rows.* In rank 4, $SO(T)_{\mathbf C}=(SL_2\times SL_2)/\mu_2$ with $T_{\mathbf C}=\mathrm{std}\boxtimes\mathrm{std}$, and the half-spin modules are $\mathrm{std}\boxtimes1$ and $1\boxtimes\mathrm{std}$.
- If $E=K$ is imaginary quadratic, then $\mathrm{Hg}(T)_{\mathbf C}=U(T,\phi)_{\mathbf C}\cong GL(W)$ with $T_{\mathbf C}=W\oplus W^*$, $\dim W=2$. This is the image of $SL_2\times\mathbf G_m$, where $\mathbf G_m$ is the diagonal torus of the second factor. The derived group $SU(T,\phi)$ is defined over $\mathbf Q$. Hence the rational splitting of $H^1(KS)$ into the part where it acts nontrivially and the part where it acts trivially is $\mathrm{Hg}$-stable. On the first part $\mathrm{Hg}$ acts through a $\mathbf Q$-form of $SL_2$ by sums of $\mathrm{std}$. On the second part it acts through a 1-dimensional torus with characters $z^{\pm1}$, so that part is isogenous to a power of a CM elliptic curve.
- If $E$ is quartic CM, then $\mathrm{Hg}(T)=U_E$ is a maximal torus of $SO(T)$, of dimension 2, and $\mathrm{Hg}(KS)$ is a maximal torus of $\mathrm{Spin}(T)$. Its characters on $H^1(KS)_{\mathbf C}$ are the four weights $\tfrac12(\pm a\pm b)$ of $S^+\oplus S^-$. So $KS$ is of CM type. A simple CM factor $B$ of dimension $g$ has $\mathrm{End}^0(B)$ a CM field of degree $2g$, whose $2g$ eigencharacters on $H^1(B)_{\mathbf C}$ restrict to distinct characters of $\mathrm{Hg}(B)$ (otherwise $\mathrm{End}_{\rm Hdg}(H^1(B))$ would be larger than $\mathrm{End}^0(B)$). Hence $2g\le4$, so $g\le2$.

`cm_rank4_weights.py` records these weight computations.

So every simple isogeny factor $B$ of some $KS(S_i)$ is one of two kinds:
- **(a)** a CM abelian variety of dimension $\le2$;
- **(b)** a non-CM factor. Its Hodge group is semisimple, its $\bar{\mathbf Q}$-simple factors are of type $A_1$ or $C_2$, all of them act nontrivially, and $H^1(B)_{\bar{\mathbf Q}}$ is a sum of standard representations of these simple factors.

Concretely, kind (b) covers:
- non-CM elliptic curves and QM surfaces (rank 3);
- abelian surfaces with RM by $F=\mathbf Q(\sqrt{\det T})$, or quaternionic fourfolds over $F$ (rank 4, $\det T\notin\mathbf Q^{\times2}$);
- two QM surfaces or elliptic curves (rank 4, $\det T\in\mathbf Q^{\times2}$);
- abelian surfaces with $\mathrm{End}=\mathbf Z$, or type II fourfolds with $\mathrm{Hg}=\mathrm{Sp}_4$-form (rank 5);
- non-CM elliptic curves and QM surfaces (rank 4, $E$ imaginary quadratic).

## 5. Hodge rings of products of the factors

**Lemma 5.1 (Goursat).** Let $B_1,\dots,B_s$ be pairwise non-isogenous simple abelian varieties of kinds (a) and (b). Then
$$\mathrm{Hg}\Bigl(\prod B_j\Bigr)=\mathrm{Hg}\Bigl(\prod_{j\in(a)}B_j\Bigr)\times\prod_{j\in(b)}\mathrm{Hg}(B_j)$$

*Proof.*
- **Semisimple part.** Let $\mathfrak h\subseteq\bigoplus\mathfrak g_j$ be the Lie algebra of the semisimple part, over $\bar{\mathbf Q}$. It projects onto every $\mathfrak g_j$.
- **A graph would appear.** If $\mathfrak h$ is not everything, the Lie-algebra form of Goursat's lemma [8] gives two simple ideals $\mathfrak s\subseteq\mathfrak g_j$ and $\mathfrak s'\subseteq\mathfrak g_k$ such that $\mathfrak h\to\mathfrak s\oplus\mathfrak s'$ is the graph of an isomorphism $\alpha$. They lie in different factors, $j\ne k$, because $\mathfrak h$ surjects onto $\mathfrak g_j$.
- **It forces an isogeny.** The standard representation is the unique irreducible representation of $\mathfrak{sl}_2$ of dimension 2, and of $\mathfrak{sp}_4$ of dimension 4. Hence $\mathrm{std}_{\mathfrak s'}\circ\alpha\cong\mathrm{std}_{\mathfrak s}$, and $\mathrm{Hom}_{\mathfrak h}(H^1(B_j),H^1(B_k))\otimes\bar{\mathbf Q}\ne0$. The rational space $\mathrm{Hom}_{\rm Hdg}(H^1(B_j),H^1(B_k))$ therefore contains a nonzero class. This is a nonzero homomorphism $B_j\to B_k$, i.e. an isogeny, which is a contradiction.
- **Mixing the two.** The kind-(a) factors have toric Hodge groups and the kind-(b) factors semisimple ones. A connected subgroup $H$ of $(\text{torus})\times(\text{semisimple})$ surjecting onto both factors is the whole product: $[H,H]$ lies in $1\times(\text{semisimple})$ and surjects onto it, and then $H$ contains it and surjects onto the torus. ∎

**Proposition 5.2.** For $B_j$ of kind (b) as in Lemma 5.1 and all $n_j\ge0$, the Hodge ring of $\prod B_j^{n_j}$ is generated by divisor classes. In particular HC holds for it.

*Proof.*
- **Reduction.** By Lemma 5.1, $H^1(\prod B_j^{n_j})_{\bar{\mathbf Q}}=\bigoplus_\alpha V_\alpha\otimes M_\alpha$, where $\alpha$ runs over the $\bar{\mathbf Q}$-simple factors $G_\alpha$ of the Hodge group, $V_\alpha$ is $\mathrm{std}$ of $SL_2$ or $\mathrm{Sp}_4$, and $M_\alpha$ is a multiplicity space. Hence
$$\Lambda^\bullet(H^1)^{\mathrm{Hg}}=\bigotimes_\alpha\Lambda^\bullet(V_\alpha\otimes M_\alpha)^{G_\alpha}.$$
- **Symplectic factors.** For $\mathrm{Sp}(V_\alpha)$ (including $SL_2=\mathrm{Sp}_2$), the first fundamental theorem says the invariants of $\otimes^kV_\alpha$ are spanned by products of contractions with the symplectic form. After antisymmetrising, $\Lambda^\bullet(V_\alpha\otimes M_\alpha)^{G_\alpha}$ is generated in degree 2.
- **Descent and conclusion.** The subalgebra generated by the rational degree-2 Hodge classes has the same dimension over $\bar{\mathbf Q}$, so the statement holds over $\mathbf Q$. Degree-2 Hodge classes are divisors by Lefschetz (1,1). ∎

**Corollary 5.3.** HC holds for every product of powers of abelian varieties whose simple factors are of kinds (a) and (b).

*Proof.* By Lemma 5.1, after grouping isogenous factors, $\mathrm{Hg}(A_{(a)}\times A_{(b)})=\mathrm{Hg}(A_{(a)})\times\mathrm{Hg}(A_{(b)})$. Hence every Hodge class is a sum of products of Hodge classes on $A_{(a)}$ and on $A_{(b)}$. The former are algebraic by Ramón-Marí's theorem for products of abelian varieties of dimension $\le2$ [1, Prop. 2.18], and the latter by Proposition 5.2. ∎

`invariants_check.py` verifies degree-2 generation by Weyl integration against explicit wedge products, for $SL_2$ on $\Lambda^{\le6}(\mathrm{std}\otimes\mathbf C^{2,3})$ and $\mathrm{Sp}_4$ on $\Lambda^{\le6}(\mathrm{std}\otimes\mathbf C^{1,2})$. It also illustrates Lemma 5.1: two isogenous $SL_2$ factors, i.e. a diagonal $SL_2$ on $\mathrm{std}\oplus\mathrm{std}$, give 3 invariants on $\Lambda^2$ against 2 for $SL_2\times SL_2$, the extra one being the mixed pairing. As a **control**, it confirms that $SL_2\times SL_2$ on $\mathrm{std}\boxtimes\mathrm{std}$, the rank-6 RM shape, has no degree-2 invariant but a degree-4 invariant on $\Lambda^4$.

## 6. Proofs of the theorems

*Proof of Theorem 1.*
- **Reduction to transcendental classes.** Let $Y=\prod S_i^{a_i}$. The Künneth components and the projectors onto $\mathrm{NS}$ and $T$ of each factor are algebraic. So $H^*(Y)$ is a sum of tensor products of pieces $H^0,H^4,\mathrm{NS}(S_i)$ (Tate type, spanned by algebraic classes) and $T(S_i)$. A Hodge class in such a summand $A\otimes W$, with $A$ algebraic of Tate type and $W=\bigotimes_iT(S_i)^{\otimes b_i}$, is a sum of products of algebraic classes with Hodge classes of $W$. It therefore suffices to show that every Hodge class $\beta\in W$ is algebraic on $\prod S_i^{b_i}$.
- **Transport to Kuga–Satake varieties.** By Corollary 2.2 and Lemma 3.1, each $T(S_i)$ has an algebraic embedding $\iota_i$ into $H^2(KS(S_i)^2)$, with an algebraic retraction $r_i$. Then $(\otimes\iota_i^{\otimes b_i})\beta$ is a Hodge class on $\prod KS(S_i)^{2b_i}$.
- **Hodge classes there are algebraic.** Every simple isogeny factor of every $KS(S_i)$ is of kind (a) or (b) (§4), so Corollary 5.3 shows the class is algebraic.
- **Back to $Y$.** Applying $\otimes r_i^{\otimes b_i}$ returns $\beta$, which is therefore algebraic. ∎

**Example 6.1 (a mixed class).** Let $E_1,E_2$ be non-isogenous elliptic curves without CM. Put $S=\mathrm{Km}(E_1\times E_2)$, with $\rho=18$ and $T(S)\cong H^1(E_1)\otimes H^1(E_2)$, and $S'=\mathrm{Km}(E_1\times E_1)$, with $\rho=19$ and $T(S')\cong\mathrm{Sym}^2H^1(E_1)$ up to twist.
- Here $\mathrm{Hom}_{\rm Hdg}(T(S'),T(S))=0$ and $\mathrm{Hom}_{\rm Hdg}(T(S'),T(S'))=\mathbf Q$.
- But $T(S)^{\otimes2}\supset\mathrm{Sym}^2H^1(E_1)\otimes\Lambda^2H^1(E_2)\cong T(S')(-1)$. So $H^6(S'\times S^2)$ contains a Hodge class not generated by classes on $S'\times S$ or $S\times S$. By Weyl integration (`example61.py`), $T(S')\otimes T(S)$ has no invariant and $T(S')\otimes T(S)^{\otimes2}$ has exactly one, while $T(S')$ itself has none. So the class is not a product of lower-degree classes.

Theorem 1 shows such classes are algebraic. In this Kummer example that is also classical, via $E_1\times E_2$. Analogous examples, for surfaces not Hodge-isogenous to any Kummer surface of a product of elliptic curves, arise from rank-4 spaces $T$ with $\det T\in\mathbf Q^{\times2}$ and $C^+(T)\cong D\times D$ for an indefinite division algebra $D$, together with the rank-3 space of trace-zero elements of $D$.

**Remark 6.2 (sharpness of the method).**
- In rank 6 with RM by a real quadratic field $E$, $H^1(KS)_{\bar{\mathbf Q}}$ contains $\mathrm{std}\boxtimes\mathrm{std}$ of $SL_2\times SL_2$. Its exterior algebra has an invariant in degree 4 but none in degree 2 (control in `invariants_check.py`).
- In rank 6 with $E=\mathbf Q$, $\mathrm{Spin}(6)=SL_4$ has the determinant invariants (Weil classes).
- So Proposition 5.2 fails beyond rank 5. There, HC needs Weil-class input (Markman [9], Floccari–Fu [10]) and, in Witt index 1, the open Kuga–Satake algebraicity discussed in [4].

## 6A. Beyond $\rho\ge17$: products in Floccari's locus

Call a projective K3 surface $S$ **of locus type** if $T(S)$ embeds isometrically in $(U^3\oplus\langle-m\rangle)_{\mathbf Q}$ for some $m>0$. By Lemma 2.1 this includes every $S$ with $\rho\ge17$. With $E=\mathbf Q$ it also includes every $S$ with $\rho=16$ and Witt index 2 [4, Thm B]. We use only $\mathrm{rk}\,T\le6$. A rank-7 $T$ would be the whole space $U^3\oplus\langle-m\rangle$, and the Kummer-type variety in [6, Lemma 5.4] would then have trivial Néron–Severi group, so we do not use that case.

**Theorem 3.** Let $S_1,\dots,S_r$ be projective K3 surfaces, each of which either has $\rho\ge17$, or is of locus type with $E(S_i)=\mathbf Q$. Then HC holds for every product $\prod S_i^{a_i}$.

The proof needs one more input.

**Lemma 6A.1.** Let $S,S'$ be of locus type and $\psi\colon T(S)\to T(S')$ a Hodge similarity. Then $\psi$ is algebraic.

*Proof.*
- By [6, Lemma 5.4] and [6, Thm 5.9], there are varieties $K,K'$ of generalised Kummer type and algebraic Hodge similarities $\alpha\colon T(S)\to T(K)$ and $\alpha'\colon T(S')\to T(K')$. These are compositions of a Buskin–Huybrechts algebraic Hodge isometry with the algebraic isometry $T(S_K)\cong T(K)(2)$ of [6, Thm 5.9].
- The map $\alpha'\psi\alpha^{-1}$ is a Hodge similarity between the transcendental lattices of two Kummer-type varieties. By Varesco [11, Thm 5.1], every Hodge morphism between Hodge-similar transcendental lattices of Kummer-type varieties is algebraic.
- **$\alpha^{-1}$ is algebraic.** Let $\dim K=2n$, $h$ ample on $K$, and $q$ the Beauville–Bogomolov form. On $T(K)\subset h^{\perp_q}$ the Fujiki relation gives $\int_Kxy\,h^{2n-2}=c\,q(x,y)$ for a nonzero constant $c$.
  - Since $\alpha$ is a similarity, $q(\alpha s,\alpha s')=\lambda\psi_S(s,s')$. So for $u\in T(K)$, $\psi_S(\alpha^{-1}u,s)=(c\lambda)^{-1}\int_Ku\,h^{2n-2}\,\alpha(s)$.
  - That is, $\alpha^{-1}=(c\lambda)^{-1}\,\pi_T\circ{}^t\!\alpha\circ(h^{2n-2}\cup\cdot)$, which is algebraic.
  - The same holds for $\alpha'$.
- Hence $\psi=\alpha'^{-1}(\alpha'\psi\alpha^{-1})\alpha$ is algebraic. ∎

Alternatively, since the Kuga–Satake correspondences of $S,S'$ are algebraic [6, Thm 5.11(i)], Varesco's theorem for K3 surfaces [11, Thm 5.3] (a consequence of [11, Cor. 4.6]) gives the lemma directly; the proof above avoids comparing the two normalisations of the Kuga–Satake correspondence.

**Lemma 6A.2 (no mixed classes between non-similar factors of rank 6).** Let $T_1,\dots,T_k$ be K3-type Hodge structures with $E=\mathbf Q$ and rank 6 (the same proof works for rank 7). Then
$$\mathrm{Hg}\Bigl(\bigoplus T_i\Bigr)=\prod_{c}SO(T_c),$$
where $c$ runs over the classes of the $T_i$ up to Hodge similarity and $SO(T_c)$ acts diagonally on the members of class $c$, through the similarities.

*Proof.*
- $\mathfrak{so}(T_i)$ is absolutely simple of type $D_3=A_3$ (or $B_3$ in rank 7).
- Every automorphism of $\mathfrak{so}_6$ preserves the vector representation up to isomorphism: the outer automorphism of $\mathfrak{sl}_4$ exchanges $4$ and $4^*$ and preserves $\Lambda^24\cong\Lambda^24^*$. $B_3$ has no outer automorphisms.
- By the Lie-algebra Goursat lemma [8], a proper subdirect product forces a graph between some $\mathfrak{so}(T_i)$ and $\mathfrak{so}(T_j)$. Then $T_j\cong T_i$ as representations of the Hodge Lie algebra, which gives a nonzero rational Hodge morphism $T_i\to T_j$.
- Since $\mathrm{End}_{\rm Hdg}=\mathbf Q$ and $(\mathrm{Sym}^2T_i)^{SO(T_i)}$ is 1-dimensional, that morphism is a similarity.
- Within a similarity class, the Hodge group is the diagonal $SO(T_c)$. ∎

*Proof of Theorem 3.*
- **Reduction.** As in the proof of Theorem 1, it suffices to treat Hodge classes in $W=\bigotimes T(S_i)^{\otimes b_i}$.
- **Splitting by blocks.** Write $W=W_1\otimes W_2$. Here $W_1$ collects the factors with $\rho\ge17$, and $W_2$ the rank-6 factors with $E=\mathbf Q$. The simple ideals of the derived Lie algebra of $\mathrm{Hg}(W_1)$ are of type $A_1$ or $C_2$ over $\bar{\mathbf Q}$ (its centre may be nontrivial, from CM factors), while $\mathrm{Hg}(W_2)$ is semisimple with simple ideals of type $D_3$. Let $\mathfrak h$ be the Lie algebra of $\mathrm{Hg}(W_1\oplus W_2)$; it surjects onto both $\mathfrak g_1$ and $\mathfrak g_2$. Then $[\mathfrak h,\mathfrak h]$ surjects onto $[\mathfrak g_1,\mathfrak g_1]$ and onto $\mathfrak g_2$, and by Goursat it is their product, because no simple ideals are isomorphic. Hence $\mathfrak h\supseteq0\oplus\mathfrak g_2$, so $\mathfrak h=\mathfrak g_1\oplus\mathfrak g_2$. Thus $\mathrm{Hg}(W_1\oplus W_2)=\mathrm{Hg}(W_1)\times\mathrm{Hg}(W_2)$ and $W^{\rm Hg}=W_1^{\rm Hg}\otimes W_2^{\rm Hg}$.
- **First block.** $W_1^{\rm Hg}$ is algebraic by Theorem 1.
- **Second block.** By Lemma 6A.2 and the first fundamental theorem for $SO$, $W_2^{\rm Hg}$ is spanned by products of two kinds of class:
  - pairings $\psi_c$ between two factors of the same class, transported by similarities;
  - determinants $\det_c\in\Lambda^{\mathrm{rk}}T_c$, likewise transported.

  The pairing classes are algebraic: they are the transcendental parts of graphs of similarities, algebraic by Lemma 6A.1 and the projectors $\pi_T$. The determinants on $S_c^{\mathrm{rk}}$ are algebraic by HC for powers of $S_c$ [6, Thm 5.11(ii)], and their transports are images under algebraic similarities. ∎

`so_controls.py` computes invariants of the diagonal and of the product groups $SO_6$, $SO_7$ by Weyl integration:
- the product $SO(T)\times SO(T')$ has no invariant on $T\otimes T'$ or on $\Lambda^3T\otimes\Lambda^3T'$;
- the diagonal (similar $T\cong T'$) has 1 invariant on $T\otimes T'$, which is the similarity;
- on $\Lambda^3\otimes\Lambda^3$ the diagonal has 2 invariants for $SO_6$ (the induced metric and the wedge into $\det$) and 1 for $SO_7$;
- $\det$ and the 3 pairings on $T^{\otimes4}$ are as predicted by the first fundamental theorem.

**Remark 6A.3 (where exceptional mixed classes come from).**
- *$E=\mathbf Q$, $\mathrm{rk}\,T\le7$.* By the arguments above, among such surfaces, mixed Hodge classes beyond similarities occur only through **split rank-4** factors ($\det T\in\mathbf Q^{\times2}$). There $SO(T)=(SL_2\times SL_2)/\mu_2$ with both factors defined over $\mathbf Q$, so one factor can be linked to a rank-3 $SO_3$ or to a factor of another split rank-4 surface (Example 6.1). Non-split rank 4 and RM factors $\mathrm{Res}_{E/\mathbf Q}SO_3$ are $\mathbf Q$-simple, so Galois-stability forbids half-graphs with them. In rank 8, triality of $D_4$ could link a vector representation to a half-spin one, and we make no claim there.
- *RM.* An RM surface of rank 6 with real quadratic $E$ has real Hodge Lie algebra $\mathfrak{sl}_2(\mathbf R)\oplus\mathfrak{so}(3)$, while a rank-4 surface has $\mathfrak{so}(2,2)\cong\mathfrak{sl}_2(\mathbf R)^2$. So no rational Goursat graph links them. (An earlier draft of this remark asserted such links; that was wrong.) For two RM surfaces with the same field, a link gives $\mathrm{Hom}_{\rm Hdg}(T,T')\ne0$. Its generator need not be a similarity, which is the twist problem of [4, §4]. For Hodge-isomorphic pairs in the locus, [4, Cor. 4.3] gives HC for $S^a\times S'^b$. We do not state more for RM.
- *Outside the locus.* Products are only as good as the powers of their factors; see [4, Rem. 4.6] for what remains.

## 7. Scope and novelty

- **Inputs from earlier work.** The rank-$\le5$ embedding (Lemma 2.1), HC for self-powers of these surfaces, and HC for products of Hodge-isomorphic pairs in the locus are proved in [4, Prop. 3.5, Cor. 3.6, Cor. 4.3], resting on Floccari's theorem [6, Thm 5.11]. Algebraicity of Hodge similarities is Varesco's [11, Thm 5.3]. HC for products of abelian varieties of dimension $\le2$ is Ramón-Marí's [1, Prop. 2.18].
- **What is new here.** The extension to arbitrary products of non-isomorphic factors: the algebraic retraction (Lemma 3.1), the classification of the Kuga–Satake factors including the two CM rank-4 cases (§4), the Goursat and invariant-theory argument (§5), and the rank-6 Goursat lemma 6A.2, which shows that similarities and determinants generate all Hodge classes on products of rank-6 locus surfaces.
- **$\rho\ge19$.** Every such K3 surface has a Shioda–Inose structure [12], so for $\rho\ge19$ Theorem 1 also follows from HC for products of elliptic curves [13]. For $\rho\in\{17,18\}$ Shioda–Inose structures exist only under lattice conditions [12]; we have not found Theorem 1 stated elsewhere.

## References

- [1] J. Ramón-Marí, On the Hodge conjecture for products of certain surfaces, Collect. Math. 59 (2008); arXiv math/0505357 (Prop. 2.18).
- [2] Yu. Zarhin, Hodge groups of K3 surfaces, J. Reine Angew. Math. 341 (1983).
- [3] M. Varesco, The Hodge conjecture for powers of K3 surfaces of Picard number 16, Michigan Math. J. 75 (2025); arXiv 2203.09778.
- [4] lelong_ridge, Algebraic real multiplication by every real quadratic field on K3 surfaces, and the Witt-index frontier in transcendental rank six, rStation hodge_conjecture_23 v4 (accepted 19 Oct 2026; Prop. 3.5, Cor. 3.6, Cor. 4.3).
- [5] J.-P. Serre, A Course in Arithmetic, GTM 7.
- [6] S. Floccari, K3 surfaces associated with varieties of generalized Kummer type, arXiv 2501.02315 (Thm 5.11, Rem. 3.2, Conj. 3.3).
- [7] P. Deligne, La conjecture de Weil pour les surfaces K3, Invent. Math. 15 (1972).
- [8] K. Ribet, Galois action on division points of abelian varieties with real multiplications, Amer. J. Math. 98 (1976) (the Lie-algebra Goursat lemma, often called Ribet's lemma).
- [9] E. Markman, Cycles on abelian 2n-folds of Weil type from secant sheaves on abelian n-folds, arXiv 2502.03415.
- [10] S. Floccari, L. Fu, The Hodge conjecture for Weil fourfolds with discriminant 1 via singular OG6-varieties, arXiv 2504.13607.
- [11] M. Varesco, Hodge similarities, algebraic classes, and Kuga–Satake varieties, Math. Z. 305 (2023); arXiv 2304.02519 (Cor. 4.6, Thm 5.3).
- [12] D. Morrison, On K3 surfaces with large Picard number, Invent. Math. 75 (1984).
- [13] H. Imai, On the Hodge groups of some abelian varieties, Kodai Math. Sem. Rep. 27 (1976).