Pith. sign in

REVIEW 4 major objections 5 minor 8 references

Higher direct images of dualizing sheaves III

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read For flat projective families whose total space has rational singularities, the higher direct images of the structure sheaf and the relative dualizing sheaf are locally free, base-change compatible, and split into their cohomology sheaves; t

desk verdict Substantial new extension of Kollár's local-freeness results to flat families with rational singularities, with a Picard scheme payoff; the main caveat is the reliance on an unstated and unproved Hodge-theoretic input from [Kol86b]. read the letter →

arxiv 2508.16507 v2 pith:SJIHMRBD submitted 2025-08-22 math.AG

János Kollár , Sándor J. Kovács This is my paper · ORCID
classification math.AG MSC 14F1714D0714B05
keywords rationalsingularitieshigherdirectimagesdualizingsheafrelativePicardschemebasechangevariationsofHodgestructuresS2extensionflatfamilies
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proves that when a flat, projective family has a total space with rational singularities, fiber cohomology cannot jump: the higher direct images of the structure sheaf and of the relative dualizing sheaf are locally free and commute with base change. It also proves that the derived direct images decompose as direct sums of their cohomology sheaves, and that the higher direct images of the absolute dualizing sheaf are obtained by tensoring the relative ones with the dualizing sheaf of the base. The main consequence is that the identity component of the relative Picard scheme is a smooth algebraic group scheme, so families of rationally singular varieties have well-behaved Jacobian-type moduli. The proof works by first establishing the claims over a large open subset where the sheaves are controlled by variations of pure Hodge structures, then extending the resulting splittings to the whole base using the S2 property of the sheaves. The revised version replaces an earlier incorrect lemma with a corrected S2-extension argument; the main theorem is unchanged.

What carries the argument

The central object is the S2-extension principle (Remark 9): if $U = S \setminus Z$ with $\operatorname{codim} Z \ge 2$ and $G$ is $S_2$, then $Hom_S(F,G) = Hom_U(F|_U,G|_U)$, the natural map $Ext^1_S(F,G) \to Ext^1_U(F|_U,G|_U)$ is injective, and a splitting $G|_U \cong G_1 \oplus G_2$ extends to a global splitting $G \cong j_*G_1 \oplus j_*G_2$. This lets the authors transfer the split surjections and split injections produced by variations of pure Hodge structures on a large open set to the entire base. The companion mechanism is Theorem 6, which packages the 1986 Hodge-theoretic input: over a large open set, all relevant direct images are locally free, determined by variations of pure Hodge structures on the smooth fibers, and the natural maps are comp

What would settle it

Construct a flat, projective morphism $g:X \to S$ over a smooth curve with $X$ having rational singularities, and a point $s$ such that $\dim H^i(X_s,O_{X_s})$ differs from the generic value. The paper's Question 14 records exactly this possibility for families with CM reduced fibers and smooth generic fiber; a positive example with rational total space would refute Theorem 1. A more targeted test is to check whether (1.4) holds in a family where $R^1g_*\omega_{X/S}$ has torsion: the theorem says it cannot, because $R^1g_*\omega_X$ is torsion-free.

Watch

Extended reading notes

Core claim

The central claim is Theorem 1: for a flat, projective morphism $g:X \to S$ of varieties over a field of characteristic zero, if $X$ has rational singularities, then (1.1) $R^i g_*O_X$ and $R^i g_*\omega_{X/S}$ are locally free and commute with base change; (1.2) $Rg_*O_X \cong \oplus_i R^i g_*O_X[-i]$; (1.3) $Rg_*\omega_{X/S} \cong \oplus_i R^i g_*\omega_{X/S}[-i]$; and (1.4) $R^i g_*\omega_X \cong \omega_S \otimes R^i g_*\omega_{X/S}$. The proof proceeds by induction on relative dimension, using a general ample divisor to reduce to lower relative dimension and using Theorem 6 (from the author's 1986 paper) to obtain split surjections and split injections over a large open subset. The key difficulty is extending these splittings from the large open set to all of $S$,

Load-bearing premise

The proof leans on a theorem from the author's 1986 paper asserting that, away from a codimension-two locus, the coherent sheaves $R^i g_*O$, $R^i g_*\omega$, and their twists are determined by variations of pure Hodge structures on the smooth fibers, and that this correspondence respects direct sums and splittings; if that Hodge-to-coherent correspondence fails to align with the coherent decomposition at the singular-fiber locus, the main theorem does not follow.

Editorial extensions

If this is right

  • For any flat, projective family over characteristic zero with rational singularities on the total space, the fiber cohomology dimensions h^i(X_s,O_{X_s}) and h^i(X_s,ω_{X_s}) are independent of s, giving deformation invariance for singular varieties.
  • The identity component of the relative Picard scheme, Pic^0(X/S), is a smooth algebraic group scheme that commutes with all base changes, so families of rationally singular varieties have well-behaved relative Jacobians.
  • For non-flat, equidimensional morphisms from a variety with rational singularities, the direct images R^i g_*O_X and R^i g_*ω_X are maximal Cohen-Macaulay, and the derived direct image decomposes into its cohomology sheaves.
  • The conclusions extend to total spaces that are Cohen-Macaulay with only a small non-rational locus: if the non-rational singularities map finitely and non-dominantly to the base, or have fiber dimension small enough, the same local freeness and Picard-scheme smoothness hold.
  • The result unifies and generalizes the previously known special case of Lagrangian fibrations, where local freeness of these direct images was established by different methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the S2-extension argument suggests the theorem should survive for base schemes that are themselves S2, provided the Hodge-theoretic input can be replaced by a suitable purity statement; the paper notes that one of the auxiliary torsion-freeness results has already been extended to general base schemes.
  • Editorial inference: the proof effectively shows that rational singularities of the total space force the coherent direct image decomposition to align with the topological Hodge decomposition on smooth fibers; a family where this alignment fails may be the right place to look for a counterexample to the theorem.
  • Editorial inference: the boundary of the class of singularities for which the theorem holds is worth probing; the paper's Example 12 shows Cohen-Macaulay is insufficient, but whether Du Bois or log canonical singularities suffice remains open.
  • Editorial inference: the smoothness of Pic^0(X/S) suggests a testable modular consequence — for any flat family of rationally singular varieties, the locus of fibers carrying nontrivial line bundles should be controlled by a smooth group scheme, so its dimension can only change through genuine base change phenomena.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proves that for a flat projective morphism g:X→S between varieties over C, if X has rational singularities, then the higher direct images R^i g_*O_X and R^i g_*ω_{X/S} are locally free and satisfy base change; moreover the derived direct images decompose as direct sums of their cohomology sheaves, and R^i g_*ω_X ≅ ω_S⊗R^i g_*ω_{X/S}. The proof proceeds by induction on relative dimension, using a large-open-subset version of the result from Kollár's 1986 papers and then extending the needed splittings across a codimension-two locus via an S2 sheaf argument. Applications include smoothness of the identity component of the relative Picard scheme and a version for non-flat equidimensional morphisms. The second version explicitly discloses and replaces an incorrect lemma from v1.

Significance. If the proof is complete, this is a strong and useful theorem: it gives deformation invariance of cohomology of the structure sheaf and the dualizing sheaf under flat deformations with rational singularities on the total space, with direct consequences for relative Picard schemes. The paper is honest about its v1 defect, includes counterexamples delimiting the result, and contains a careful treatment of dualizing sheaves in Sections 16–23. However, the central proof rests on a Hodge-theoretic input from [Kol86b,p.174] that is not precisely stated or proved; until that input is made explicit and verifiable, the main theorem is conditional.

major comments (4)
  1. [Theorem 6, proof] The splitting of δ°: g°_*ω_{H°/S°}→R^1g°_*ω_{X°/S°} is the linchpin of the proof of (1.1). It is supplied entirely by the quotation from [Kol86b,p.174] asserting that the coherent direct images are 'determined by' the VHS of the smooth fibers and that this correspondence 'respects direct sums'. The authors themselves state that the cited theorem is 'not stated explicitly enough' for their purposes. Semisimplicity of VHS gives a splitting over the smooth locus, but the extension of that splitting to the large open set S°—where the discriminant may be a divisor—is exactly what needs proof. Please state the precise theorem used, with hypotheses, and either prove it or give a citation that literally covers the maps in (6.2). Without this, R^1g_*ω_{X/S} is not shown to be locally free.
  2. [Paragraph 8, after (8.3)] The proof says 'It is thus enough to prove the claims when Y is smooth' after replacing Y by a resolution r:Y'→Y. This is not immediate for the maps in (6.2): the sheaves O_H, ω_H, and the restriction/Gysin maps are not invariant under an arbitrary resolution unless the divisor H is sufficiently transverse to the exceptional locus. In Paragraph 8, H is only a 'general, sufficiently ample' divisor, and it may meet Sing(X) in positive dimension. Please spell out the comparison of the maps involving H, or arrange H to be disjoint from the singular locus in a way compatible with the resolution.
  3. [Corollary 5, proof] The sentence 'g_*ω_{X/S}(H) is locally free (since ω_{X/S} is flat over S)' is not justified by flatness alone: flatness of a sheaf does not imply that its zeroth direct image is locally free. This property usually follows if H is chosen sufficiently ample so that R^i g_*(ω_{X/S}(H))=0 for all i>0 and the base-change theorem applies. Please add the required vanishing/ampleness condition or a reference.
  4. [Corollary 5, proof] The step 'π_*R^i g_*O_X = R^i g'_*O_X is locally free, so R^i g_*O_X is CM' requires justification. For a finite morphism π, local freeness of π_*F does not in general imply F is CM. For example, let A=k[[v,z]] and A'=k[[z^2,v]]; then A is finite flat over A' and the maximal ideal (v,z) of A is free as an A'-module, yet it has depth one over A. If the intended argument uses S2/reflexivity of R^i g_*O_X, that property must be proved first; as written, the proof of Corollary 5 is incomplete.
minor comments (5)
  1. [Theorem 1, (1.2)] The symbol 'P' in (1.2), (1.3), (5.2), (5.3), and (6.3) appears to mean the direct sum; please define it or use \bigoplus.
  2. [Remark 7] The displayed isomorphism '⊕^d2_1 L1' is hard to parse; please clarify the intended direct sum and the subscript/superscript conventions.
  3. [Paragraph 9] Typo: 'Let F, Gbe coherent sheaves' is missing a space.
  4. [Paragraph 23 heading] Typo: 'V erifying Assumption 8.4' should be 'Verifying Assumption 8.4'.
  5. [Theorem 24] The hypotheses on Z should explicitly state the codimension/non-dominance condition needed to choose a general very ample H disjoint from Z; as written, 'Z→S is finite, non-dominant' is not enough when the relative dimension is one.

Circularity Check

0 steps flagged · score 2.0 of 10

Heavy reliance on the same authors' earlier large-open-subset theorems, but the flat-everywhere conclusion is not assumed in the cited work; no circular reduction is exhibited.

full rationale

Theorem 1's proof is not circular by the hard-rule standard. Its main new content is the upgrade from a large-open-subset statement to all of S: the paper explicitly says, 'As we discuss in Theorem 6, [Kol86b] implies that (1.1–4) hold over a large open subset S^0 ⊂ S (even if g is not assumed flat). Thus the main assertion is that if g is flat, then these hold everywhere.' The cited [Kol86a, Kol86b] results are used for the weaker open-subset version, local freeness over S^0, and torsion-freeness; the flat-everywhere conclusion is not among their assumptions. The step that closes the gap, the S2-extension in Remark 9, is original and was corrected in v2 after Bhatt's objection, confirming it is not an imported artifact. Assumption 8.4 is flagged as an assumption and then proved in Paragraph 23 using Lemma 22 and Corollary 19, neither of which assumes (1.1)–(1.4). The only genuinely load-bearing self-citation is the 'main theorem of [Kol86b, p.174]' used in the proof of Theorem 6, which asserts that the relevant direct images are determined by variations of pure Hodge structures and that the correspondence respects direct sums. The paper candidly states that this theorem is 'not stated explicitly enough for our current purposes.' That is a rigor gap and a heavy reliance on the first author's earlier work, but it is not circular: the cited theorem is an external, previously published statement with assumptions that do not include the target theorem, and no fitted parameter or prediction-equivalent-to-input is present. No equation of the paper reduces to an input by construction, and no 'prediction' is a renamed fit. Thus the circularity score is low; the weaknesses are about dependence on an unstated Hodge-theoretic theorem, not about self-definitional reasoning.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The free-parameter and invented-entity ledgers are empty: this is a pure mathematics paper with no fitted constants and no newly postulated objects. The assumptions are standard theorems and domain hypotheses; the most fragile is the non-explicit [Kol86b] Hodge correspondence.

assumptions (5)
  • standard math Variations of pure Hodge structures form a semi-simple category; every morphism is a composite of a split surjection and a split injection (Deligne, [Del71]).
    Used in the proof of Theorem 6 to obtain the splitting statements (6.2) over a large open subset of S.
  • standard math The main theorem of [Kol86b, p.174]: on a large open subset the sheaves R^i g_* O, R^i g_* O(-H), R^i g_* ω, R^i g_* ω(H) are determined by variations of Hodge structures on smooth fibers and the correspondence respects direct sums.
    This is the load-bearing input for Theorem 6; the paper says the source theorem is not stated explicitly in [Kol86b] and does not restate it.
  • standard math Rational singularities in characteristic 0 are Cohen-Macaulay; ω_X is S2; and a flat morphism from a variety with rational singularities has base S with rational singularities (Boutot).
    Used in Paragraph 8 and Remark 2 to justify the S2 extension argument and the base change statements.
  • domain assumption A general sufficiently ample divisor H on X has rational singularities and g|_H is flat.
    Used for the induction step in Paragraph 8; this is a standard Bertini-type statement not proved here.
  • standard math R^i g_* ω_X is torsion-free for proper morphisms to varieties ([Kol86a, 2.1]).
    Used to rule out a torsion summand in the verification of Assumption 8.4 and in Theorem 24.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Higher direct images of dualizing sheaves III." pith.science (2026). https://pith.science/paper/SJIHMRBD

@misc{pith2026250816507,
  author       = {Pith},
  title        = {Pith review of: Higher direct images of dualizing sheaves III},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJIHMRBD}},
  note         = {Machine review of arXiv:2508.16507}
}
read the original abstract

We show that for flat morphisms between varieties with rational singularities, the higher direct images of the structure sheaf are locally free. As a consequence, the identity component of the relative Picard scheme is a smooth algebraic group scheme. v.2. Bhatt pointed out that Lemma 9 was incorrect; it is now replaced. The main theorem is unchanged.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages

  1. [1]

    I , Global Analysis (Papers in Honor of K

    [Art69] Michael Artin, Algebraization of formal moduli. I , Global Analysis (Papers in Honor of K. Kodaira), Univ. Tokyo Press, Tokyo, 1969, pp. 21–71. MR MR0260746 (41 #5369) [BLR90] Siegfried Bosch, Werner L¨ utkebohmert, and Michel Raynaud, N´ eron models, Ergeb- nisse der Mathematik und ihrer Grenzgebiete (3), vol. 21, Springer-Verlag, Berlin,

    work page 1969

  2. [6]

    [KK20] J´ anos Koll´ ar and S´ andor J Kov´ acs,Deformations of log canonical and F -pure singulari- ties, Algebr. Geom. 7 (2020), no. 6, 758–780. MR 4156425 [Kol86a] J´ anos Koll´ ar,Higher direct images of dualizing sheaves. I , Ann. of Math. (2) 123 (1986), no. 1, 11–42. MR 825838 (87c:14038) [Kol86b] , Higher direct images of dualizing sheaves. II , A...

    work page Pith review arXiv 2020

  3. [8]

    [Mat05] Daisuke Matsushita, Higher direct images of dualizing sheaves of Lagrangian fibrations , Amer. J. Math. 127 (2005), no. 2, 243–259. MR 2130616 [Mur25] Takumi Murayama, Relative vanishing theorems for Q-schemes, Algebr. Geom. 12 (2025), no. 1, 84–144. MR 4841227 [SGA70] Sch´ emas en groupes. I: Propri´ et´ es g´ en´ erales des sch´ emas en groupes,...

    work page 2005

  4. [1962]

    MR MR0146040 (26 #3566) [Kim24] Yoon-Joo Kim, The N´ eron model of a higher-dimensional Lagrangian fibration, https: //arxiv.org/abs/2410.21193,

    arXiv

  5. [1990]

    Bombieri and D

    MR 1045822 (91i:14034) [BM76] E. Bombieri and D. Mumford, Enriques’ classification of surfaces in char. p. III, Invent. Math. 35 (1976), 197–232. MR 491720 [Bou87] Jean-Fran¸ cois Boutot, Singularit´ es rationnelles et quotients par les groupes r´ eductifs, Invent. Math. 88 (1987), no. 1, 65–68. MR MR877006 (88a:14005) [Con00] Brian Conrad, Grothendieck d...

    work page 1976

  6. [2000]

    MR MR1804902 (2002d:14025) [DBJ74] Philippe Du Bois and Pierre Jarraud, Une propri´ et´ e de commutation au changement de base des images directes sup´ erieures du faisceau structural, C. R. Acad. Sci. Paris S´ er. A 279 (1974), 745–747. MR 0376678 (51 #12853) [Del71] Pierre Deligne, Th´ eorie de Hodge. II, Inst. Hautes ´Etudes Sci. Publ. Math. (1971), no...

    work page arXiv 1974

  7. [2024]

    [Extraits du S´ eminaire Bourbaki, 1957–1962.] , Secr´ etariat math´ ematique, Paris,

    HIGHER DIRECT IMAGES OF DUALIZING SHEA VES III 11 [Gro62] Alexander Grothendieck, Fondements de la g´ eom´ etrie alg´ ebrique. [Extraits du S´ eminaire Bourbaki, 1957–1962.] , Secr´ etariat math´ ematique, Paris,

    work page 1957

  8. [2025]

    N\'eron models, minimal models, and birational group actions

    [Kol25b] , N´ eron models, minimal models, and birational group actions, https://arxiv. org/abs/2502.13800,

    work page Pith review arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.