Minimal Model Program¶

A birational-classification program that follows canonical-divisor-negative extremal directions through contractions and flips, seeking a model with nef log canonical divisor or a Mori fiber-space endpoint while admitting only controlled singularities.

Version: v1 · 2026-08-30 · History
Domain-specific #: 2277
Origin domain: birational algebraic geometry
Subdomain: Mori theory
Aliases: Mori program, Mori's program, Log minimal model program, MMP

Core Idea¶

The minimal model program (MMP), also called the Mori program, is the organizing procedure for birationally classifying projective algebraic varieties by replacing a given variety or pair with a simpler birational model. Its directional instrument is the canonical divisor: for a log pair \((X,\Delta)\), the program examines the numerical behavior of \(K_X+\Delta\) on curves. If that divisor is nef, meaning \((K_X+\Delta)\cdot C\geq 0\) for every curve class \(C\), the pair has reached the minimal-model side of the program. If it is not nef, the cone theorem identifies a \((K_X+\Delta)\)-negative extremal ray, and the contraction theorem supplies a morphism that contracts exactly the curves in that direction.^[1]^[2]

The contraction selects one of three structural moves. A divisorial contraction removes a divisor and lowers the Picard number. A small contraction removes no divisor but cannot simply be retained in the working category; the program seeks its flip, replacing the contracted negative curves by curves on which the transformed log canonical divisor is positive. A fiber-type contraction ends the run with a Mori fiber space rather than another birational model of equal dimension. The procedure repeats after a divisorial contraction or flip.^[3]^[4]

This is a program rather than a universally terminating algorithm. Its steps depend on hypotheses concerning projectivity, characteristic, divisor classes, and singularities; an extremal ray need not be unique; and termination of arbitrary higher-dimensional flip sequences is not known in full generality. Major theorems establish the program in important regimes. In particular, Birkar, Cascini, Hacon, and McKernan prove existence and finite-generation results for projective Kawamata log terminal pairs under bigness hypotheses, including minimal and canonical models for smooth projective varieties of general type.^[3] The stable abstraction is therefore not the false claim that every variety mechanically reduces to a unique smooth object. It is the controlled, canonical-divisor-directed birational classification architecture and its two characteristic endpoint types.

Structural Signature¶

projective Q-factorial log pair (X,Delta) in an admitted singularity class -> test nefness of K_X+Delta -> if non-nef choose a negative extremal ray R -> contract R -> divisorial replacement, flip, or fiber-type exit -> repeat while preserving birational data and singularity control -> nef minimal model or Mori fiber space, subject to existence and termination hypotheses

The roles are:

Input pair. Usually a normal projective variety together with a boundary divisor, placed in a class such as klt or dlt for which discrepancies and the needed theorems are meaningful.
Directional divisor. The log canonical divisor \(K_X+\Delta\) determines which curve directions are negative.
Cone of curves. Numerical curve classes form \(\overline{NE}(X)\); a negative extremal ray is a one-dimensional boundary direction eligible for contraction.
Extremal contraction. A contraction morphism collapses the curves whose numerical classes lie in the chosen ray.
Move classifier. The relative dimensions and exceptional locus distinguish divisorial, small, and fiber-type contractions.
Flip when required. A small contraction is replaced by another small birational model over the same base, with the sign of the log canonical divisor reversed on the contracted locus.
Controlled singularities. Smoothness is generally not preserved in dimension at least three; the working singularity category must be stable enough under the moves.
Progress and endpoint. Divisorial contractions reduce relative Picard number. Flips require a separate termination argument. A run ends with \(K_X+\Delta\) nef or with a Mori fiber space.

Birational equivalence is the global invariant: the same function field is retained through the birational steps. Smoothness, individual curve classes, and the isomorphism type are deliberately allowed to change.

What It Is Not¶

Not a generic “minimal model.” Many fields minimize parameter counts, description length, loss, or mechanism complexity. The MMP is fixed to birational algebraic geometry, the log canonical divisor, extremal rays, contractions, and flips.
Not numerical model selection. AIC, cross-validation, regularization, and sparse regression compare statistical models under predictive or inferential criteria. They do not operate in a birational class or test nefness of \(K_X+\Delta\).
Not one theorem. Cone, contraction, base-point-free, flip-existence, termination, abundance, and finite-generation results play different roles. Their conjunction supports versions of the program, but “the MMP” is the framework organizing them.
Not resolution of singularities. Resolution replaces a singular variety by a smoother birational one. The MMP commonly travels in the opposite direction, accepting controlled singularities to obtain a numerically simpler model.
Not a canonical-form algorithm. A minimal model need not be unique. Distinct minimal models can be related by flops, so equality of outputs is not the recognition condition.
Not the canonical model. A minimal model has nef \(K_X+\Delta\); a canonical or log canonical model has an ample log canonical divisor in the relevant sense and is typically obtained from a finitely generated canonical ring. Nefness alone does not make a model canonical.
Not guaranteed termination in arbitrary generality. A draft that says “repeat until done” without naming the theorem or assumptions converts a research program into an unsupported universal algorithm.

Scope of Application¶

The home domain is birational algebraic geometry. The program classifies complex projective varieties and log pairs, studies Fano and general-type geometry, analyzes moduli spaces through their birational models, and structures questions about canonical rings and singularities. The language also extends to relative settings over a base, to MMP with scaling, and—under separate theorem packages—to some positive-characteristic, mixed-characteristic, and compact Kähler settings. Those extensions preserve the core roles but do not erase their different hypotheses.^[5]^[6]

The program genuinely recurs across dimensions and object classes rather than naming one publication. On smooth projective surfaces it recovers classical contraction of \((-1)\)-curves. In dimension three it requires the new mechanisms of flips and controlled singularities, established through Mori theory. In higher dimensions, the log-pair formulation and finite generation results provide working programs in major cases such as varieties of general type.^[1]^[6]^[3] “MMP for foliations” and other newer variants are qualified adaptations because their ambient objects and theorem statements change; they are not evidence that the unqualified node covers every geometric simplification scheme.

Clarity¶

A compact recognition test is: What is the divisor that directs the process, what negative extremal ray is being contracted, which contraction case occurs, what singularity class is preserved, and which theorem licenses the next step or endpoint? If the answers are \(K_X+\Delta\), an extremal ray of the cone of curves, divisorial/small/fiber type, an explicit category such as Q-factorial klt pairs, and an applicable existence or termination result, the MMP is probably present. If the answers are merely “make the model simpler,” the label is probably generic.

The phrase “minimal” is numerical, not synonymous with smallest dimension, fewest equations, least singular, or unique. A minimal model can be singular and can have the same dimension as the starting variety. “Program” signals a family of linked conjectures, theorems, and permitted moves, not pseudocode with an unconditional halt. “Model” means a birational representative of the same function field, not a statistical prediction device.

Manages Complexity¶

Birational equivalence classes contain many varieties and many possible birational maps. Directly comparing all representatives would be intractable. The MMP compresses this search by turning the canonical divisor into a compass. Rather than enumerate arbitrary birational modifications, the geometer asks where \(K_X+\Delta\) is negative, uses the cone and contraction theorems to isolate a permitted extremal direction, and applies one of a small number of move types. This converts a sprawling classification problem into a controlled path through the cone of curves.

The abstraction also separates local and global obligations. The contraction is localized on curves in one extremal ray, but the output must remain in an admissible global singularity class. Divisorial steps carry an elementary progress certificate because the Picard number decreases. Flips preserve that number, exposing termination as an independent burden rather than allowing “complexity decreased” to remain a slogan. Finally, the two endpoint forms organize the classification: pseudo-effective canonical direction points toward minimal models, while the non-pseudo-effective side leads toward Mori fiber geometry under suitable hypotheses.^[3]

Abstract Reasoning¶

The MMP licenses a disciplined sequence of inferences:

If \(K_X+\Delta\) is nef, no \((K_X+\Delta)\)-negative extremal contraction remains, so the run has reached the minimal-model criterion used by the program.
If it is not nef and the cone/contraction hypotheses hold, a negative extremal ray supplies an admissible next direction; arbitrary contractions are not interchangeable with this one.
If the contraction is divisorial, relative Picard number falls, so only finitely many such steps can occur in a fixed finite-rank setting.
If it is small, the target generally fails to retain the divisor behavior needed to continue; a flip is the prescribed replacement, and flip existence must be established.
If it is fiber type, the output is a Mori fiber space and the program has classified the input on the negative-canonical side rather than failed to find a minimal model.
If different ray choices produce different minimal models, that nonuniqueness is compatible with the program; one studies their relation, often through flops, instead of declaring one run incorrect.
If a proposed generalization changes the canonical divisor, singularity category, or permitted surgery, its theorems must be re-established rather than inherited by vocabulary.

These moves distinguish a proof of one run from a claim about all runs. Existence of a flip is not termination of flip sequences; existence of one minimal model is not uniqueness; finite generation in a stated regime is not the abundance conjecture.

Knowledge Transfer¶

The exact mechanism transfers within algebraic geometry. For surfaces, threefolds, log pairs, relative morphisms, and selected characteristic regimes, researchers map the same roles—log canonical divisor, negative ray, contraction type, singularity control, endpoint—onto different theorem packages. MMP with scaling adds an auxiliary ample divisor and varies a scaling parameter so that the next wall or extremal ray becomes controlled, but it remains an MMP because the canonical-divisor-directed contraction logic persists.^[2]

Outside birational geometry, only a structural analogy transfers: preserve an equivalence class, identify a directed defect, apply local simplifications, and stop at a normal form or structured residual. Compiler optimization, graph reduction, and statistical model selection may share that skeleton, yet they lack canonical divisors, extremal rays, discrepancies, and Mori fiber spaces. Those domain-bound objects carry the mathematical content. The transferable residue is already represented by broad primes such as Transformation and Equivalence-Preserving Rewriting; it does not make the MMP itself a prime.

Examples¶

Blowing down an exceptional curve on a surface. Let \(X=\operatorname{Bl}_p\mathbf{P}^2\), the blow-up of the projective plane at one point, and let \(E\) be the exceptional curve. Then \(E\cong\mathbf{P}^1\), \(E^2=-1\), and \(K_X\cdot E=-1\). Its numerical class spans a canonical-negative extremal ray. The contraction sends \(E\) to the original smooth point and returns \(\mathbf{P}^2\). This is a divisorial contraction: the exceptional locus has codimension one, the Picard number drops from two to one, and no flip is needed. Classical surface minimal-model theory repeatedly performs this move until no eligible \((-1)\)-curve remains, unless the geometry exits through a ruled or fiber-type case.^[4]

A small contraction in dimension three. In higher dimension an extremal contraction can have exceptional locus of codimension at least two. Simply replacing the source by the contraction target may make the canonical divisor insufficiently Cartier for intersection theory. A flip replaces the small contraction \(X\to Z\) by \(X^+\to Z\): the two models are isomorphic away from the exceptional loci, while \(K_X+\Delta\) is negative over \(Z\) and \(K_{X^+}+\Delta^+\) is positive. This example exhibits why smooth-surface intuition is incomplete and why controlled singularities are part of the identity.^[1]^[6]

General type. For a smooth projective variety of general type over characteristic zero, BCHM proves finite generation of the canonical ring and the existence of minimal and canonical models. A run may reach a log terminal model with nef canonical divisor; the associated canonical model is obtained by the canonical ring and has a stronger ampleness property. The example keeps minimal and canonical outputs distinct while showing how the program becomes a theorem in an important regime.^[3]

Structural Tensions¶

Simplification versus singularity. Contracting geometrically negative directions simplifies the birational model but can create singularities. Requiring smooth outputs would block necessary higher-dimensional steps; allowing arbitrary singularities would destroy discrepancy and intersection control. Diagnostic: name the singularity class and verify it is preserved by the proposed move.

Choice versus classification. Several negative extremal rays may be available. Choosing one makes the run executable but can produce a different minimal model or Mori fiber structure. Diagnostic: distinguish invariants of the birational class from artifacts of the chosen run, and do not infer uniqueness from termination.

Local progress versus global termination. Each flip improves the sign of the canonical divisor near one contraction, but the Picard number need not fall. Local improvement therefore does not itself rule out an infinite sequence. Diagnostic: cite an actual termination theorem or keep termination conditional.

Minimal versus canonical. Nefness supplies a flexible endpoint compatible with flops; ampleness and the canonical-ring construction supply a more rigid canonical model when available. Diagnostic: test whether the claim uses nef or ample, and whether finite generation or abundance has been invoked.

Birational preservation versus geometric loss. Function field and birational class persist, but divisors, curves, smoothness, and other geometric data may change. Diagnostic: state the exact preserved relation rather than saying the variety is “the same.”

Structural–Framed Character¶

The Minimal Model Program is predominantly structural (aggregate 0.1). Its input, ordering criterion, moves, invariants, and endpoints are mathematically specified: divisor intersection, extremal rays, morphisms, discrepancies, nefness, and ampleness. “Simple” and “minimal” can sound evaluative, but within the program they are resolved into numerical and categorical conditions rather than aesthetic judgment.

The residual framing lies in the research-program boundary. Different variants choose base field, pair type, singularity class, and endpoint theorem, and the historical phrase “Mori program” groups results and conjectures whose reach developed over time. That does not make the mechanism institutional or conventional; it requires readers to treat every universal-sounding claim as hypothesis-indexed.

Structural Core vs. Domain Accent¶

The portable core is equivalence-preserving directed reduction: hold a coarse identity fixed, use a directional invariant to select an obstruction, replace the object by a controlled equivalent form, and iterate toward an endpoint. This core supports general reasoning about progress certificates, local-versus-global termination, and nonunique normal forms.

The domain accent is constitutive: normal projective varieties or log pairs; birational equivalence; \(K_X+\Delta\); numerical curve classes and the Mori cone; extremal rays; divisorial, small, and fiber-type contractions; flips and flops; discrepancy-controlled singularities; nef minimal models; and Mori fiber spaces. Remove these and one no longer has the Minimal Model Program. The domain package is why the candidate survives composite closure by general Transformation, Iteration, Optimization, Canonical Form, and Equivalence-Preserving Rewriting.

The Minimal Model Program presupposes Transformation: every contraction or flip is a rule-governed geometric replacement with explicit invariants and permitted changes. The proposed DAG therefore uses a single strict composition edge to prime:transformation. Subsumption would be misleading because the MMP is an iterative research framework containing several transformations, not itself one mapping.

It is closely related to Equivalence-Preserving Rewriting, because the steps navigate among birational models while preserving the function field and use canonical-divisor negativity to direct choices. It is not a strict instance of that prime's full signature in every formulation: the prime requires a separate cost ranking over an admissible rewrite space, whereas an MMP run can choose among several extremal rays without a total or scalar cost order. Iteration captures repetition but not legal moves. Canonical Form is an important contrast because minimal models need not be unique. Classification names the goal but omits the machinery.

Relationships to Other Abstractions¶

Current abstraction Minimal Model Program Domain-specific

Parents (1) — more general patterns this builds on

Minimal Model Program presupposes Transformation Prime

The Minimal Model Program presupposes Transformation: every contraction or flip is a rule-governed geometric replacement with explicit invariants and permitted changes.

Hierarchy path (1) — routes to 1 parentless root

Minimal Model Program → Transformation → Function (Mapping)

Neighborhood in Abstraction Space¶

Minimal Model Program sits in a sparse region of the domain-specific corpus (84^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Minimal model: an endpoint object, not the whole program that seeks it.
Canonical model: a stronger ample/canonical-ring endpoint, not every nef minimal model.
Mori fiber space: the characteristic fiber-type endpoint on the negative side, not a failed or incomplete run.
Mori theory: the broader body of extremal-ray and contraction theory on which the MMP is built; often used nearly synonymously in informal context but not an exact object-level identity.
Minimal model theory in logic: the study of models with minimal embedding properties, unrelated despite the words “minimal model.”
Model selection: a statistical comparison procedure, not birational classification.
Resolution of singularities: a smoothening birational modification, whereas the MMP may introduce controlled singularities.
Flip: one move required after a small contraction, not the entire program.
Flop: a crepant birational move, often relating different minimal models; it does not eliminate a canonical-negative ray in the same way as a flip.
Sarkisov program: a framework decomposing birational maps between Mori fiber spaces; it uses MMP outputs but has a different target relation.

References¶

[1] Kollár, János, and Shigefumi Mori. Birational Geometry of Algebraic Varieties. Cambridge Tracts in Mathematics 134. Cambridge University Press, 1998. https://doi.org/10.1017/CBO9780511662560. Standard authoritative treatment of rational curves, the cone and contraction machinery, log pairs, singularities, flips, and the general program. registry ↩a ↩b ↩c

[2] Corti, Alessio, Anne-Sophie Kaloghiros, and Vladimir Lazić. “Introduction to the Minimal Model Program and the Existence of Flips.” Bulletin of the London Mathematical Society 43, no. 3 (2011): 415–448. https://doi.org/10.1112/blms/bdq126; arXiv:0811.1047. Specialist exposition of Mori-theoretic foundations, log singularities, models, scaling, and conditional flip construction. registry ↩a ↩b

[3] Birkar, Caucher, Paolo Cascini, Christopher D. Hacon, and James McKernan. “Existence of Minimal Models for Varieties of Log General Type.” Journal of the American Mathematical Society 23, no. 2 (2010): 405–468. https://doi.org/10.1090/S0894-0347-09-00649-3. Primary source for finite generation and minimal/log canonical model existence under stated hypotheses. registry ↩a ↩b ↩c ↩d ↩e

[4] Matsuki, Kenji. Introduction to the Mori Program. Universitext. Springer, 2002. https://doi.org/10.1007/978-1-4757-5602-9. Authoritative textbook covering surface birational geometry, logarithmic pairs, contractions, flips, and the program's workflow. registry ↩a ↩b

[5] Tanaka, Hiromu. “Minimal Model Program for Excellent Surfaces.” Annales de l'Institut Fourier 68, no. 1 (2018): 345–376. https://doi.org/10.5802/aif.3163. Primary example of a theorem package extending the program to an explicitly qualified setting. registry ↩

[6] Mori, Shigefumi. “Flip Theorem and the Existence of Minimal Models for 3-Folds.” Journal of the American Mathematical Society 1, no. 1 (1988): 117–253. https://doi.org/10.2307/1990969. Primary threefold theorem establishing the role of flips and minimal models. registry ↩a ↩b ↩c

[7] Kollár, János. “The Structure of Algebraic Threefolds: An Introduction to Mori's Program.” Bulletin of the American Mathematical Society 17, no. 2 (1987): 211–273. https://doi.org/10.1090/S0273-0979-1987-15548-0. Authoritative survey of the program's classification architecture and threefold context. registry