Can flips go on forever?

0:00Does every flip sequence stop?

The minimal model program simplifies a variety by a sequence of surgeries, guided by an adjoint divisor, K + B: the canonical divisor plus a boundary. At each stage, pick an extremal ray, a class of curves on which K + B is negative, and contract it. If the contraction swallows a divisor, that divisor is gone. If it collapses something smaller, the program performs a flip: cut out that locus, and glue in a different one, on which K + B becomes positive. The program should end where K + B is nef, never negative on any curve, or at a Mori fibre space, a fibration along which K + B stays negative. Divisorial contractions cannot go on forever, since each one lowers a count of divisor classes, a whole number that cannot fall forever. But a flip leaves that count unchanged, so why should flips stop? In dimension three, Shokurov’s counting device, the difficulty, gave the first termination proofs. In dimension four, answers came case by case: Kawamata, Matsuda and Matsuki for terminal singularities, Fujino for canonical pairs, Alexeev, Hacon and Kawamata when the negative of K + B is numerically equivalent to an effective divisor, Chen and Tsakanikas, and Moraga, when K + B is pseudo-effective, and Han, Liu and Zhuang for klt pairs with big boundary. Shokurov and Birkar showed that well-chosen programs reach an end. What remained was every program, for log canonical fourfolds, without such positivity assumptions.

1:36The claims

Result 056 bundles five OpenAI preprints. The projective one, 44 pages, dated September 24, claims the four-dimensional case of termination of flips for log canonical pairs with rational boundary, over any algebraically closed field of characteristic zero. Its theorem has three parts. Every negative extremal ray, at every stage, has its contraction and, when that is birational, its next model. Every program has finitely many steps, for every sequence of choices, with no pseudo-effectivity and no Q-factoriality assumed, and mixed steps, a divisorial contraction followed by a small modification, are allowed. And every maximal program ends at a nef adjoint or a Mori fibre space. The other four work on compact Kähler fourfolds, which need not be projective. A 47-page paper, dated October 7, claims that every sequence of generalized log canonical flips terminates, with rational boundary, fixed nef data, and flips given by projective small diagrams, on globally Weil Q-factorial fourfolds. It concerns flips that exist; it does not claim every flip exists. A 22-page companion treats the generalized canonical case. And an 89-page paper, dated October 6, claims that for klt pairs with rational boundary, every maximal ordinary program terminates, at a nef model or a projective Mori fibre space. Its September version claimed only that some finite program exists.

3:11Counting what improves

Here is how the projective proof goes. Compare the two sides of a step on a common model. Log discrepancies, which measure how mild the singularities are along each divisor over the variety, never decrease, and they strictly increase exactly over the bad locus, where either map fails to be an isomorphism. The bad loci on the two sides have dimensions adding up to at least three. Ranks of cycle classes show that only finitely many steps lose a divisor, so an infinite program ends in a tail of flips. Once surfaces stop appearing on the new side, a second rank count finishes the job. So everything comes down to one question: can surfaces keep appearing on the new side forever? To answer it, the paper builds an integer called the difficulty, on auxiliary terminal models over the steps. It weighs each divisor exceptional over these models by how far its log discrepancy sits below two, scaled by a fixed integer N and rounded up. Along a typical surface inside a boundary component of coefficient b, there is a ladder of echoes, with log discrepancies 1 + j(1 − b). With b equal to 0.7 and N equal to ten, the echoes sit at 1.3, 1.6, and 1.9, with weights seven, four, and one, a default of twelve at every such surface. Following Alexeev, Hacon and Kawamata, the difficulty subtracts this default surface by surface, and adds back divisor cycle ranks on the normalized boundary.

4:45The branch inequality

The new trouble is that some coefficients on these models keep decreasing, so N cannot clear them. Where several branches of the normalized boundary meet one surface, rounding can fall short. With N equal to ten and coefficient 0.11, three branches round to four together, but to six separately. The shortfall is at most the number of varying branches minus one. The key geometric input, the branch inequality, bounds the total excess of branches by cycle ranks on the normalizations, plus divisors centered on curves with log discrepancy below two, both already counted in the difficulty. So at the terminal model over each step, the difficulty is never negative. The proof starts in topology: mixed Hodge theory and the decomposition theorem isolate local systems along singular curves, which become line bundles, and then distinct divisors. Two more facts finish the klt case. Lowering coefficients and running small flips never increase the difficulty: across a flip, the change in the defaults cancels exactly against the change in the cycle ranks. And each surface on the new side forces a drop of at least one, because there the scaled amount below two is an integer, so any strict improvement crosses a rounding threshold. With N equal to ten, a log discrepancy rising from 1.69 to 1.7 lowers its weight from four to three. A nonnegative integer that never increases can drop only finitely often. Special termination extends this to log canonical pairs, and a transfer argument carries it from the complex numbers to every algebraically closed field of characteristic zero.

6:30Where it stands

All five are preprints, not yet peer reviewed, and there is no Lean formalization. The projective paper has not been revised; three of the Kähler papers were replaced in the October 7 update, after versions dated October 5. The finite-program paper and the log canonical flip paper cite each other, and both explain why this is not circular. The projective paper cites three other release papers, including log abundance from episode 034, but only for consequences at the endpoints. We found no public assessment. If they hold, every program, for every choice of extremal rays, on a projective log canonical fourfold with rational boundary in characteristic zero must stop, and on compact Kähler fourfolds, so must every ordinary program for klt pairs with rational boundary, under the stated factoriality assumptions. The general case with real coefficients, positive characteristic, and higher dimensions remain beyond these papers.