Family 056 — Why four-dimensional minimal model programs should stop

← All digests

Family 056 · Algebraic and complex geometry
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-12

1. The question

An algorithm that repeatedly improves an object is useful only if it eventually stops. In algebraic geometry, the minimal model program simplifies a space by contracting curves along which its canonical geometry is negative. Sometimes it removes a divisor. Sometimes it performs a flip, replacing a small exceptional set without reducing the number of divisors.

That second operation makes termination difficult. Local improvement of singularities does not immediately provide a positive integer that decreases at every step. An infinite sequence could, in principle, make ever smaller improvements.

Family 056 contains five manuscripts about four-dimensional spaces. They claim termination in broad projective and compact Kähler settings, with careful distinctions between constructing one finite program, terminating every permitted program, and terminating flip sequences whose individual maps are already supplied. These are different assertions even when their eventual geometric pictures look alike.

2. Background and essential definitions

Pairs and positivity

A pair $(X,B)$ consists of a normal space and an effective boundary $B=\sum b_jB_j$, a weighted sum of hypersurfaces. The adjoint divisor is $D=K_X+B$, where $K_X$ records top-degree differential forms. Rational boundary means all $b_j$ are rational numbers.

A divisor is nef if its class is a limit of positive classes; in projective geometry this is equivalent to nonnegative intersection with every curve. It is pseudo-effective if its class is a limit of effective divisor classes. In the Kähler setting the manuscripts use the corresponding definition through positive currents, allowing differential-form classes not represented by divisors.

If a curve direction has $D$-degree negative, the program seeks to contract it. A divisorial contraction removes a hypersurface. A small contraction removes no hypersurface; its exceptional locus has higher codimension. A flip replaces a small contraction $X\to Z$ by $X^+\to Z$ so that $-D$ is relatively ample on the first side and $D^+$ is relatively ample on the second.

A Mori fiber space is the other possible endpoint: a contraction to a lower-dimensional base, with connected fibers, relative Picard number one, and relatively ample negative adjoint. Relative Picard number one means there is only one independent numerical direction for divisors along the fibers.

Measuring singularities

On a resolution $p:W\to X$, the log discrepancy of a prime divisor $E$ is

$$a(E;X,B)=1+\operatorname{coeff}_E\bigl(K_W-p^*(K_X+B)\bigr).$$

Log canonical, abbreviated lc, means every such number is nonnegative. Klt means every one is positive. Canonical and terminal impose, respectively, at least one and greater than one for exceptional divisors, with the additional positivity conditions specified in the manuscripts.

A generalized pair also carries a nef divisor on a higher birational model. Its compatible traces form a nef b-divisor $M$, and the adjoint becomes $K_X+B+M_X$. Fixing this higher-model datum throughout a sequence is part of the problem; choosing unrelated nef terms at each step would be a different operation.

Projective versus Kähler

A projective variety embeds in projective space. A compact Kähler space has a compatible positive closed differential form and can be nonprojective. A morphism between nonprojective spaces may nevertheless be projective because it has a relatively ample line bundle.

The Kähler manuscripts require global Weil $\mathbb Q$-factoriality: every prime divisor defined on the whole space has a positive Cartier multiple, and the canonical sheaf has a positive invertible reflexive power. This is not a claim about every analytic local ring. Their stronger global condition applies to every rank-one reflexive sheaf. These distinctions determine which line bundles and contractions are available.

3. Development history and previous partial results

Three-dimensional minimal model theory established the fundamental pattern of contractions, flips, and endpoints. In higher dimensions, Birkar–Cascini–Hacon–McKernan supplied major existence results for klt pairs with positivity assumptions. Existence of flips and existence of selected minimal models did not, by themselves, prove that every possible sequence terminates. BCHM.

Shokurov introduced difficulty, a weighted count of exceptional divisors with small discrepancies. Kawamata–Matsuda–Matsuki combined discrepancy counts with surface-cycle ranks for terminal fourfolds. Fujino extended the approach to canonical fourfold pairs, with an essential correction treating repeated blowups over boundary surfaces. Alexeev–Hacon–Kawamata developed further weighted corrections and obtained termination under a numerical effectivity assumption on the negative adjoint. Fujino’s addendum, Alexeev–Hacon–Kawamata.

For pseudo-effective fourfolds, Chen–Tsakanikas and Moraga proved broad termination results, including generalized lc settings. Han–Liu–Zhuang subsequently treated every klt fourfold program with big boundary. The gap addressed by the projective manuscript is rational-boundary lc termination without either pseudo-effectivity or a big boundary. Non-pseudo-effectivity does not imply that the negative adjoint is numerically effective, so the older assumptions cannot simply be exchanged. Chen–Tsakanikas, Han–Liu–Zhuang.

In compact Kähler geometry, Höring–Peternell developed threefold minimal models. Das–Hacon–Păun obtained fourfold results when the adjoint is rationally equivalent to an effective divisor. Fujino’s analytic relative theory supplies projective operations, while Hacon–Xie’s 2026 work gives generalized cone and base-point-free results and programs in big settings. These are important ingredients and partial solutions, rather than consequences newly created by this family. Fujino’s analytic MMP, Hacon–Xie.

4. What the new manuscripts claim

M1–M2: termination of supplied Kähler flip diagrams

M1 claims that every sequence of generalized lc flips on globally Weil $\mathbb Q$-factorial compact Kähler fourfolds terminates, subject to its precise data. The boundary is effective and rational. The fixed nef b-divisor is represented by an analytically nef $\mathbb Q$-Cartier divisor on a projective modification. The adjoints are actual $\mathbb Q$-Cartier divisors with compatible canonical representatives.

Each supplied flip diagram has projective small maps with connected fibers to a normal compact Kähler base and the opposite ample signs described above. No scaling rule or pseudo-effectivity is required, and the bases may vary. The theorem concerns these existing diagrams; it does not itself construct every requested contraction or flip.

M2 gives the generalized-canonical case with all discrepancies positive and exceptional ones at least one. Boundary coefficients are less than one, and discrepancy exactly one is permitted. It supplies a more direct discrepancy-and-cycle argument within that narrower singularity class. Its maps also need not have relative Picard number one.

M3: arbitrary projective lc programs

M3 claims that, over every algebraically closed field of characteristic zero, every permitted program for a projective lc fourfold with rational boundary terminates. Neither initial $\mathbb Q$-factoriality nor pseudo-effectivity is assumed.

Its allowed steps include divisorial contractions, flips, and a divisorial contraction followed by a small positive modification. The negative side contracts one extremal ray. Arbitrary blowups, flops, and inserted isomorphisms are not allowed steps. It also claims existence of the required continuation for every selected negative ray.

Every maximal program therefore ends with nef adjoint or a Mori fiber space. A maximal program means one continued until an endpoint is reached, not one stopped by choice halfway through.

M3 separately obtains semiample endpoints using the abundance manuscript in family 034, and local complement and Cartier-section consequences using companion results in families 066 and 068. Those additions use extra theorems; termination alone does not imply them.

M4–M5: ordinary programs starting on the given Kähler space

M4, the October 6 manuscript, claims termination of every maximal ordinary negative-ray program from a globally Weil $\mathbb Q$-factorial compact Kähler klt fourfold with effective rational boundary. The initial model is exactly the given space. Its contractions are projective, and its endpoint has nef adjoint if the initial adjoint is pseudo-effective; otherwise it is a Mori fiber space over a compact Kähler base.

The conclusion includes the corresponding intrinsic formulation using the canonical rational line bundle, rather than requiring a chosen global canonical divisor. Strong global factoriality, if present initially, persists. No abundance conclusion is included.

M5, the September 24 manuscript with the same title, proves the weaker quantifier: there exists one finite ordinary program with these endpoint alternatives. Its limiting-model construction is a separate route. It should not be described as an additional arbitrary-termination theorem merely because it also produces a finite result.

The family thus contains related constructions with different scopes, rather than five interchangeable formulations of one assertion.

5. Why the advance matters

Importance: 5/5. Termination is a central structural problem in birational geometry. If correct, these claims would remove major remaining restrictions in dimension four and extend the available picture substantially into nonprojective Kähler geometry.

The projective result strengthens “a suitable path reaches an endpoint” to “every allowed path does.” The analytic results also address the existence and compatibility of operations when no global ample polarization is available. Together, these would make four-dimensional classification less dependent on special choices or positivity assumptions.

The endpoint distinction is equally consequential. A nef canonical adjoint is a minimal-model outcome. Producing sections of its multiples is the further abundance problem. Keeping those stages separate reveals exactly which part of the classification program each theorem would settle.

6. The main idea

A corrected integer that can only decrease

Flips improve discrepancies, but infinitely many valuations can arise from blowing up ordinary boundary surfaces. Counting all of them naively gives an unusable infinite quantity. The difficulty method subtracts the predictable families and retains the exceptional contributions.

M1 and M3 pass to auxiliary terminal models, called junctions, carrying the relevant low-discrepancy divisors. Coefficients can still vary along the sequence. The manuscripts confine persistent variation to divisors mapping to curves or points, where a new branch estimate controls the correction terms.

A boundary component can have several branches over a surface after normalization, which separates local branches into a normal space. If there are $r$ branches, rounding their individual weights can exceed rounding the total by as much as $r-1$. The elementary inequality is

$$\left\lceil\sum_{j=1}^r x_j\right\rceil \ge \sum_{j=1}^r\lceil x_j\rceil-(r-1).$$

The geometric work pays for this deficit using cycle ranks on normalized components and additional divisorial contributions over curves. The resulting corrected difficulty is nonnegative and drops by at least one whenever a relevant flipped surface appears. Topological decomposition and Hodge-theoretic tools enter precisely to control the branch excess.

After those positive-side surfaces disappear, remaining flips remove independent surface-cycle classes on the negative side. A finite-dimensional rank cannot decrease forever. These two decreasing quantities address different stages of a potential infinite sequence.

Canonical witnesses and construction of steps

M2 uses a more direct witness: blow up the smooth general point of a flipped surface. With fixed rational boundary and nef data, its discrepancy lies in a finite set between one and two. Strict improvement prevents repeated use indefinitely. Cycle-rank descent handles the remaining steps. Smoothness is established where the witnesses are needed; it is not assumed everywhere on a canonical space.

M4 adds supporting contractions and flip algebras for ordinary negative rays, then invokes M1 for the existing flip tail. M1 in turn uses relative constructions, extraction, and special termination developed in M4. The sources describe those lower-level results as independent inputs to the final deductions; the two headline theorems should not be portrayed as independent proofs of one another.

M5 instead follows a limiting-model argument with Kähler scaling. A hypothetical infinite sequence yields a limiting terminal pair. Rational nef decompositions and exact descent of line bundles preserve a positive intersection gap; a small perturbation then contradicts the negativity of an earlier ray. Its auxiliary extractions do not replace the initial model of the ordinary program.

7. Scope, limitations, and open questions

The principal statements are four-dimensional. They do not settle arbitrary termination in every dimension. Rational boundaries and fixed rational nef data are essential stated hypotheses of the relevant theorems, even when auxiliary arguments use real classes.

A theorem about existing generalized lc flip diagrams is not an unrestricted existence theorem for generalized lc Kähler programs. The ordinary construction theorem requires klt singularities and global factoriality. Conversely, M3’s projective lc theorem has no initial factoriality requirement. Transferring assumptions between these statements would alter their scope.

None provides a numerical upper bound for the number of steps from simple input data. Nor does finiteness make endpoints unique. Nefness does not by itself imply semiampleness, and an analytic space carrying projective contractions need not become projective over a point.

Questions beyond the stated claims include higher-dimensional termination, broader real or transcendental generalized data, effective bounds, and abundance at the resulting analytic endpoints.

8. Sources and further reading

M1 and M2 state their termination results in Theorem 1.1. M3’s Definition 1.1 and Theorem 1.2 specify exactly which programs and mixed steps are permitted. M4 and M5 both use Theorem 1.2, with the crucial difference between every maximal program and existence of one finite program. Their introductions and dependency discussions guide the strategy above.

The primary historical papers linked in Section 3 explain difficulty, corrected boundary counts, pseudo-effective termination, and the analytic construction of steps. The constituent source links below retain the October editions where listed, as well as the distinct September finite-program manuscript.

Source manuscripts

• M1. OpenAI, Termination of generalized log canonical flips on compact Kähler fourfolds (October 7, 2026).
• M2. OpenAI, Termination of generalized-canonical flips on compact Kähler fourfolds (October 6, 2026).
• M3. OpenAI, Termination for projective log canonical fourfolds with rational boundary (September 24, 2026).
• M4. OpenAI, Finite ordinary minimal model programs on compact Kähler fourfolds (October 6, 2026).
• M5. OpenAI, Finite ordinary minimal model programs on compact Kähler fourfolds (September 24, 2026).

Sources consulted on 12 October 2026 at repository revision fd4aeeb. Download the manuscript citations.