How Do Geometrically Irreducible Perverse Sheaves Spread Out Over Arithmetic Bases?

What spreading out actually produces

The paper's central object is a geometrically irreducible perverse sheaf M on the generic fiber X_k of a finite-type morphism X_0 to S_0, where S_0 is an integral scheme of finite type over Z[1/ℓ] with function field k. The main theorem states that there is a dense open subscheme U_0 of S_0 and a sheaf M_0 in Perv_ULA(X_0/U_0) whose restriction to the generic fiber is isomorphic to M. Two properties matter here. The extension is relatively perverse over U_0, meaning perversity is measured relative to the base rather than absolutely on each fiber. It is also universally locally acyclic, so its formation behaves well under base change.

The uniqueness clause is conditional: if U_0 is geometrically unibranch, the extension to U_0 is unique up to unique isomorphism. That condition is not decoration. Without it, the construction only pins down the extension after passing to an étale neighborhood and twisting by a rank-one sheaf coming from that neighborhood, which is exactly the mechanism the paper uses to make the generic fiber generically unramified.

Why a twist is needed before extending

The obstruction to globalizing the earlier trait-based equivalence of categories is that a perverse sheaf on a generic fiber need not be generically unramified. The paper's main observation is that geometric irreducibility repairs this after an étale localization: over an étale neighborhood of the base, the sheaf becomes generically unramified once it is twisted by a rank-one sheaf coming from that neighborhood. An arithmetic character is always generically unramified by the quasi-unipotent monodromy theorem, which is why the twist can be absorbed when the base is arithmetic.

This is the difference between the arithmetic and the general regular case. Over an integral arithmetic base, the theorem extends the sheaf itself. Over a regular base that is not arithmetic, the paper obtains an extension of a twist over a dominant étale neighborhood instead. Readers comparing the two statements should keep that distinction: the untwisted extension is an arithmetic-base phenomenon, not a general-base one.

From lisse sheaves to perverse sheaves

The extension theorem is deduced from a variant for geometrically irreducible lisse sheaves, proved separately for the arithmetic case and for the general case. The reduction passes through a relative intermediate extension: given an immersion j: V_0 to X_0 and a relatively perverse sheaf M on V_0, the paper constructs a dense open U_0 of S_0 and a relatively perverse, universally locally acyclic sheaf N on X_{U_0}/U_0 whose fiber at each point x of U_0 equals j_{!*}(M_x). Generic base change is what allows the construction to commute with base change over a dense open, and a further shrinking makes N universally locally acyclic.

A second ingredient handles the geometry of the fibers. If the reduced generic geometric fiber of X_0 over S_0 is smooth, then after a universal homeomorphism of the base there is a dense open U_0 over which the reduced fibers are smooth. In characteristic zero this is immediate; in characteristic p the argument uses a Frobenius twist to reach a geometrically reduced fiber first. This is the step that lets the paper control fundamental groups fiber by fiber.

Decomposition series that survive in families

For the applications, the paper needs more than an extension of one irreducible sheaf: it needs the decomposition series of a relatively perverse sheaf to persist across the family. Theorem 1.17 provides this over a topologically Noetherian integral geometrically unibranch base S_0 with function field k, assuming the irreducible subquotients of the generic fiber have geometrically irreducible support. After shrinking the base, there is a filtration 0 = M_0 ⊂ M_1 ⊂ ... ⊂ M_n = M_{U_0} by relatively perverse, universally locally acyclic sheaves, together with closed subsets Z_i of X_{U_0} and dense open V_i of Z_i.

The listed properties are what make the filtration usable: the generic fiber of each quotient M_i/M_{i−1} is irreducible; each Z_i has geometrically irreducible fibers over U_0; each V_i has geometrically connected fibers with smooth reduced geometric fibers; each quotient restricts to a lisse sheaf on V_i; the image of the fundamental group in the general linear group of a fiber is independent of the chosen geometric point up to conjugacy; and each fiber of the quotient is the intermediate extension of its restriction to V_i. The geometric-irreducibility hypothesis on supports is what makes the monodromy specialization theorem applicable, since it guarantees that dense opens of the supports stay geometrically connected.

Specializing to finite fields to reach purity

The applications all follow one template. Start with a geometrically irreducible perverse sheaf over an arithmetic field, spread it out to an arithmetic base scheme, and shrink the base so that the fiber over every closed point is geometrically irreducible. Lafforgue's theorem then implies that the fiber over a closed point is pure up to a twist. That purity is what the theory of weights needs in order to make statements about nearby cycles.

This is why the paper can move from a statement over an arithmetic field to a statement over a finite field without losing control. The spreading-out theorem supplies the model; the closed-point fibers supply the finite-field setting; purity supplies the weight-theoretic input. The paper notes that the resulting theory is softer than the earlier trait-based equivalence of categories in one respect: after spreading out, specialization is just taking a fiber over a point of the base, rather than applying an equivalence of categories and then taking the fiber at the closed point.

What the arithmetic Kashiwara conjecture gains

The first application is the arithmetic Kashiwara conjecture of Esnault and Kerz, proved in equicharacteristic for traits of geometric origin. Let S be the spectrum of the strict henselization of a discrete valuation ring essentially of finite type over a field, let X be a scheme of finite type over S, and let M be a geometrically semisimple arithmetic perverse sheaf on the generic fiber X_η. The theorem has two parts. First, the graded pieces Gr^M_n(Ψ_uni(M)) of the monodromy filtration on the unipotent nearby cycles are geometrically semisimple for every integer n. Second, for a proper morphism f: X to Y of finite type S-schemes, the monodromy spectral sequence with E^pq_1 term the perverse cohomology of f_*(Gr^M_{−q}(Ψ_uni(M))) degenerates on the second page, and the filtration it induces agrees with the monodromy filtration.

The scope qualifiers carry weight. The result is stated in equicharacteristic and for traits of geometric origin, not for arbitrary traits. The input sheaf is geometrically semisimple and arithmetic, not an arbitrary perverse sheaf. And the conclusion is about the graded pieces of the monodromy filtration and about degeneration on the second page, not about the full nearby cycles being semisimple.

Decomposition and Hard Lefschetz over arithmetic fields

The second application recovers the decomposition theorem for arithmetic complexes by specializing to the known finite-field case through the models built earlier. The paper then generalizes it: if k is a field finitely generated over a separably closed field, f: X to Y is a proper morphism of finite type k-schemes, and K is an arithmetic semisimple complex on X, then the pushforward f_*(K) is semisimple. The authors state that they expect the method to extend to complexes over a scheme in characteristic zero without the arithmeticity assumption, by appealing to the decomposition theorem for semisimple complexes over an algebraically closed field of characteristic zero, but that extension is presented as an expectation rather than a proved theorem.

The third application is a Hard Lefschetz statement. For a field k, a projective morphism f: X to Y of finite type k-schemes, the Chern class η in H^2(X, Q_ℓ(1)) of an f-ample line bundle on X, and a geometrically semisimple arithmetic perverse sheaf M on X, the morphism η^i from ^p H^{−i}(f_*(M)) to ^p H^i(f_*(M))(i) is an isomorphism for all i ≥ 0. Here the hypotheses are projective rather than proper, and the sheaf is geometrically semisimple and arithmetic.

Where the argument stops

The generic base change results in the appendix are the technical backbone of the specialization arguments and are flagged as of independent interest. For a finitely presented morphism f: X to Y of qcqs schemes, with coefficients either in a complete Noetherian local ring whose residue characteristic is invertible on Y or in Q_ℓ for a prime ℓ invertible on Y, there is a finite constructible partition of Y into locally closed pieces on which the pushforward is constructible and commutes with base change, the complex is universally locally acyclic, and an internal mapping object exists and commutes with base change. The paper is explicit that the last two items are not proved in the stated generality for complete Noetherian I-adic coefficient rings, because of technical points the authors could not resolve; the precise statements are left to the appendix.

The practical reading is therefore a split by hypothesis rather than a single uniform statement. Untwisted extension over a dense open requires an arithmetic base; over a general regular base one gets an extension of a twist over a dominant étale neighborhood. The Kashiwara application requires equicharacteristic traits of geometric origin and a geometrically semisimple arithmetic input. The decomposition theorem is proved for arithmetic semisimple complexes over fields finitely generated over a separably closed field, with the characteristic-zero non-arithmetic case left as an expectation. The Hard Lefschetz statement requires a projective morphism and a geometrically semisimple arithmetic perverse sheaf. The most useful thing a reader can carry away is this mapping from hypothesis to conclusion: geometric irreducibility of the input is what buys the extension, arithmeticity of the base is what removes the twist, and specialization to closed points is what converts the extension into purity and hence into the weight-theoretic conclusions.

Sources used in this answer