Disproof of the Yau--Tian--Donaldson conjecture

Abstract: We construct a polarized smooth projective fivefold and prove that it is K-polystable but does not admit a constant scalar curvature Kähler metric. This disproves the Yau--Tian--Donaldson conjecture for constant scalar curvature metrics. The main result of this paper was obtained using generative AI, particularly GPT-5.6-sol, Fable 5, and the Danus system. A detailed report on the use of generative AI in this paper is enclosed in the appendix, joint with Bin Dong and Guoxiong Gao.

• The paper constructs an explicit smooth polarized fivefold that is K-polystable at every positive exponent while its polarization admits neither an extremal nor a constant-scalar-curvature Kähler metric.
• The paper encodes the obstruction in an admissible boundary polynomial with an irrational double zero, yielding approximate cscK metrics with Calabi-functional infimum zero but no genuine extremal metric.
• The paper classifies all zero-Donaldson–Futaki normal test configurations as products and exhibits rational degenerations with vanishing stability ratios, separating ordinary K-polystability from uniform K-stability.

Central claim and scope

The paper claims to disprove the constant-scalar-curvature form of the Yau–Tian–Donaldson (YTD) conjecture by constructing an explicit smooth polarized projective fivefold $(X,A)$ that is K-polystable with respect to every normal ample algebraic test configuration, at every positive exponent, but whose polarization class $c_1(A)$ contains no extremal Kähler metric and therefore no cscK metric (2608.19301).

The result concerns the ordinary Donaldson–Futaki notion of K-polystability for general polarized varieties. It does not challenge the established YTD correspondence for smooth Fano manifolds, nor the stronger formulations involving uniform, reduced, completed, or quantitatively coercive K-stability. Indeed, a central feature of the example is that it is K-polystable but not uniformly relatively K-polystable.

The construction addresses a specific gap in earlier candidate counterexamples arising from admissible projective bundles. Previous work produced polarized varieties for which the admissible extremal polynomial is nonnegative on rational points but has an irrational interior zero. Such a polynomial obstructs an extremal metric, but proving K-semistability or K-polystability requires controlling arbitrary normal test configurations rather than only the explicitly constructed admissible degenerations. The present paper’s principal contribution is the claimed classification of all zero-Donaldson–Futaki test configurations for the chosen example.

The geometric construction

Let $C_0,C_1,C_2,C_3$ be smooth projective curves with genera

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$

chosen so that their Jacobians are pairwise Hom-orthogonal:

$$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$

The paper obtains such curves by taking hyperelliptic curves with sufficiently general Galois groups. Zarhin’s theorem gives absolute simplicity of each Jacobian, while the distinct genera exclude isogenies between different factors. The product

$B=C_0\times C_1\times C_2\times C_3$

is therefore a smooth projective fourfold with highly constrained endomorphism and automorphism structure.

Choose line bundles $M_i,L_i$ with degrees

$\deg M_i=(461999,13962,1068,260)$

and

$\deg L_i=(13397971,-11635,-712,-104).$

Set

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$

and define

$c_1(A)$0

Since $c_1(A)$1 has dimension four and the fibers of $c_1(A)$2 are projective lines, $c_1(A)$3 is a smooth projective fivefold. The endpoint degree vectors are

$c_1(A)$4

and

$c_1(A)$5

all of whose entries are positive. This gives ampleness of $c_1(A)$6. The automorphism calculation yields

$c_1(A)$7

with the connected automorphism group generated by scaling the $c_1(A)$8-summand in $c_1(A)$9. The absence of vector fields on the high-genus base and the vanishings $C_0,C_1,C_2,C_3$0 are essential here. In particular, all connected automorphisms arise from fiber scaling.

The boundary polynomial and the analytic obstruction

The admissible geometry reduces the relevant scalar-curvature and stability calculations to a one-variable polynomial. Define

$C_0,C_1,C_2,C_3$1

and

$C_0,C_1,C_2,C_3$2

With the chosen numerical data, the factors can be written as

$C_0,C_1,C_2,C_3$3

$C_0,C_1,C_2,C_3$4

$C_0,C_1,C_2,C_3$5

$C_0,C_1,C_2,C_3$6

All four factors are positive on $C_0,C_1,C_2,C_3$7. Let $C_0,C_1,C_2,C_3$8 be determined by

$C_0,C_1,C_2,C_3$9

with endpoint conditions

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$0

The paper computes the polynomial exactly:

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$1

Thus $(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$2 is nonnegative on $(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$3, vanishes at the endpoints, and has a unique interior zero

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$4

of multiplicity two. The irrationality and double-contact at $(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$5 are the decisive features of the construction.

For admissible metrics with profile $(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$6, the normalized scalar curvature is

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$7

Perturbing $(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$8 by

$(g_0,g_1,g_2,g_3)=(3846511,10591,76,46),$9

produces smooth admissible Kähler metrics $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$0 satisfying

$$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$1

Consequently, the Calabi functional has infimum zero even though the limiting profile is not strictly positive. This is important: the example does not exhibit a positive lower bound for scalar-curvature deviation. Instead, it produces approximate cscK metrics whose degeneration concentrates at the irrational double zero.

The nonexistence proof proceeds by contradiction. Suppose $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$2 contained an extremal metric. Maximal compactness implies that the fiber-scaling circle acts isometrically. Extremal metrics persist under small deformations of the Kähler class, while uniqueness identifies nearby extremal metrics with the explicit admissible extremal metrics associated with nearby profiles. Their momentum coordinates converge uniformly, and their profiles converge to

$$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$3

whose unique interior zero occurs at $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$4. On each projective-line fiber, the limiting momentum coordinate assumes every value between $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$5 and $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$6, so some interior point has momentum $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$7. At that point the fiber-scaling vector field is nonzero, but the limiting identity

$$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$8

would force $$\operatorname{Hom}\bigl(\operatorname{Pic}^0(C_i),\operatorname{Pic}^0(C_j)\bigr)=0 \qquad (i\neq j).$$9. This contradicts positive definiteness. Hence $B=C_0\times C_1\times C_2\times C_3$0 contains no extremal metric and, a fortiori, no cscK metric.

K-semistability from approximate metrics

The approximate scalar-curvature result is converted into K-semistability using Donaldson’s lower bound for the Calabi functional. For a normal ample test configuration with positive centered norm, the Donaldson–Futaki invariant provides a lower bound for the scalar-curvature error of every Kähler metric. Since the infimum of that error in $B=C_0\times C_1\times C_2\times C_3$1 is zero for every $B=C_0\times C_1\times C_2\times C_3$2, a negative Donaldson–Futaki invariant is impossible.

The scaling relation is explicit. If $B=C_0\times C_1\times C_2\times C_3$3, then $B=C_0\times C_1\times C_2\times C_3$4 and, because $B=C_0\times C_1\times C_2\times C_3$5,

$B=C_0\times C_1\times C_2\times C_3$6

Thus every normal ample test configuration for $B=C_0\times C_1\times C_2\times C_3$7 has nonnegative Donaldson–Futaki invariant. The zero-norm case is handled separately using the characterization of zero non-Archimedean norm by almost triviality, which already forces vanishing of the Donaldson–Futaki invariant.

This proves K-semistability at every positive exponent, but not yet K-polystability. The difficult issue is the equality case: a zero Donaldson–Futaki invariant could, a priori, arise from a nonproduct degeneration.

Classification of zero-invariant degenerations

The classification argument is the technical core of the paper. It combines filtrations, Newton–Okounkov convex transforms, Chow stability of the base blocks, rationality properties of Duistermaat–Heckman measures, normalization under base change, and relative Smith normal forms.

For a test configuration $B=C_0\times C_1\times C_2\times C_3$8 with $B=C_0\times C_1\times C_2\times C_3$9, the section spaces decompose under fiber scaling as

$M_i,L_i$0

Taking the two opposite fiber-scaling initials of the filtration produces block-preserving filtrations. Their convex transforms are analyzed on the Newton–Okounkov body

$M_i,L_i$1

The argument first proves that the slice average of either convex transform has the form

$M_i,L_i$2

The only possible nonlinearity is therefore a crease at the irrational point $M_i,L_i$3. Rational slices are treated using multiplication maps among the spaces

$M_i,L_i$4

The paper proves exponentially accurate asymptotic scalarity for the partial traces of multiplication projectors on the curve factors. Tensorization over the four curves then yields a fixed-pair inequality for blockwise mean filtration weights. Together with Chow polystability of the high-degree product embeddings of $M_i,L_i$5, this implies that the asymptotic Chow invariant vanishes on every rational slice. A strictness theorem for filtrations over cscK polarized manifolds then forces the restriction of the transform to each rational slice to be constant.

Convexity propagates this constancy to every interior slice, leaving only the scalar profile $M_i,L_i$6. The irrational crease is then excluded by the algebraicity of the Duistermaat–Heckman measure. Finitely generated algebraic filtrations have piecewise polynomial weight densities with rational coefficients and rational breakpoints. The pushforward of the affine pieces of $M_i,L_i$7 would produce a density whose spectral data detects the location of the crease. Since $M_i,L_i$8 is irrational, this is impossible. The paper treats both monotone and sign-changing slope configurations, including the equal-height case in which the two branches meet at the same image point. In the latter case, the coefficient involves

$M_i,L_i$9

which is irrational, contradicting rationality of the algebraic density.

It follows that both opposite initials have affine transforms

$\deg M_i=(461999,13962,1068,260)$0

The common affine coefficients are then shown to be rational. The argument uses rationality of all moments of an algebraic Duistermaat–Heckman measure and the explicitly computed first three central moments of the density

$\deg M_i=(461999,13962,1068,260)$1

The nonzero variance and third central moment eliminate the ambiguity between slopes $\deg M_i=(461999,13962,1068,260)$2 and $\deg M_i=(461999,13962,1068,260)$3, while rational second and third moments imply $\deg M_i=(461999,13962,1068,260)$4 and then $\deg M_i=(461999,13962,1068,260)$5.

From affine transforms to product test configurations

After clearing denominators by a ramified base change $\deg M_i=(461999,13962,1068,260)$6, the affine profile becomes integral. The paper constructs a product comparator $\deg M_i=(461999,13962,1068,260)$7 whose increasing filtration jump on $\deg M_i=(461999,13962,1068,260)$8 is

$\deg M_i=(461999,13962,1068,260)$9

where

$\deg L_i=(13397971,-11635,-712,-104).$0

are integers.

The normalized base change of $\deg L_i=(13397971,-11635,-712,-104).$1 remains zero-invariant. This uses a comparison between a test configuration and its normalization: normalization can only decrease the Donaldson–Futaki invariant by a nonnegative term determined by the leading coefficient of the Hilbert polynomial of the normalization quotient. Since semistability gives nonnegativity after normalization and ordinary base change preserves the zero invariant, the normalization quotient must contribute no leading term. The normalized and unnormalized configurations consequently have the same Duistermaat–Heckman law.

The residual filtration is obtained by subtracting the product profile from the filtration of $\deg L_i=(13397971,-11635,-712,-104).$2. Its convex transform is zero, so its normalized jump measures converge to $\deg L_i=(13397971,-11635,-712,-104).$3. Since the variety has dimension five, the paper derives the quantitative estimate

$\deg L_i=(13397971,-11635,-712,-104).$4

for the relative Smith elementary divisors $\deg L_i=(13397971,-11635,-712,-104).$5 between the marked lattices of $\deg L_i=(13397971,-11635,-712,-104).$6 and $\deg L_i=(13397971,-11635,-712,-104).$7. The exponent $\deg L_i=(13397971,-11635,-712,-104).$8 is precisely $\deg L_i=(13397971,-11635,-712,-104).$9.

A marked rigidity proposition shows that this subquadratic Smith growth forces equality of the marked section algebras. The proof is divisorial: if the normal section algebras differed, a height-one valuation would generate a subspace of dimension order $M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$0 on which the lattice separation is order $M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$1, forcing a Smith-square sum of order $M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$2, contradicting the little-$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$3 estimate.

Finally, descent through the deck group introduces a ceiling:

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$4

The summability estimate for row defects shows that the total ceiling error over each block row remains uniformly bounded. A rational arithmetic lemma then forces

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$5

so the ceiling disappears identically. The original test configuration therefore has exactly the affine product filtration on every block. This proves that every zero-invariant normal ample test configuration is a polarized product induced by fiber scaling and a scalar character.

Separation from uniform K-stability

The example is K-polystable but has no positive uniform stability margin. The paper constructs rational single-crease test configurations with crease parameters

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$6

where $M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$7 is the Fibonacci sequence. These satisfy

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$8

The associated test configurations have positive Donaldson–Futaki invariant

$M=\boxtimes_i M_i,\qquad L=\boxtimes_i L_i,$9

where

$c_1(A)$00

Thus $c_1(A)$01 decays cubically in $c_1(A)$02. By contrast, the reduced non-Archimedean $c_1(A)$03-functional remains uniformly bounded below after normalization:

$c_1(A)$04

Consequently,

$c_1(A)$05

This explicitly separates ordinary K-polystability from uniform relative K-polystability. The irrational double zero is invisible to every individual rational algebraic crease, which has strictly positive Donaldson–Futaki invariant, but it is approached by rational creases whose invariants tend to zero.

Scheme-theoretic and definitional consequences

The paper extends the nonnegativity statement beyond normal test configurations. For an arbitrary ample test configuration, normalization gives a normal configuration with no larger Donaldson–Futaki invariant. Hence the fivefold remains semistable for arbitrary ample scheme-theoretic test configurations. Equality holds precisely when the normalization is the relevant product and the normalization morphism is an isomorphism away from a subset of codimension at least two.

This distinction is necessary. Under the unrepaired convention that permits equality only for literal products, even $c_1(A)$06 would fail the equality condition despite carrying the Fubini–Study cscK metric. The paper constructs a nonnormal test configuration with product normalization, zero Donaldson–Futaki invariant, and a codimension-two defect. This confirms that codimension-two triviality, rather than literal scheme-theoretic productness, is the appropriate equality notion.

Limitations and open questions

The counterexample has an infinite connected automorphism group:

$c_1(A)$07

The paper therefore leaves open whether the ordinary cscK YTD conjecture might hold under the additional assumption that $c_1(A)$08 is trivial or finite. The construction also relies on a four-dimensional base and extremely large genera and degrees. The fourth curve factor is not merely decorative: its Jacobian orthogonality and the resulting multiplication and rigidity properties are used to control arbitrary test configurations. Whether an analogous argument can produce a lower-dimensional example, or a smooth example with finite automorphism group, is not established.

The claims are also unusually dependent on a long chain of analytic and algebraic inputs: admissible extremal theory, openness and uniqueness of extremal metrics, filtered Okounkov-body asymptotics, Chow stability, rational spectral structure, normalization comparisons, and marked Smith rigidity. The paper explicitly reports extensive use of generative AI in discovering and developing the argument (2608.19301). That provenance does not alter the mathematical assertions, but it makes independent verification of the universal classification theorem particularly important: the decisive statement quantifies over all normal ample test configurations and is not reducible to a finite numerical certificate.

Conclusion

The paper presents a claimed explicit smooth polarized fivefold for which ordinary Donaldson–Futaki K-polystability does not imply cscK existence. Its construction encodes an irrational double zero into an admissible boundary polynomial, producing approximate cscK metrics but obstructing extremal metrics. The main technical achievement is the claimed classification of every zero-invariant normal test configuration as a product, achieved through affine convex transforms, rational Duistermaat–Heckman constraints, base change, and marked Smith-spectrum rigidity. The resulting example separates ordinary K-polystability from uniform K-stability and leaves the finite-automorphism case as a specific unresolved question.