This webpage is about Lean 3, which is effectively obsolete; the community has migrated to Lean 4.

Jump to the corresponding page on the main Lean 4 website.

Lean Glossary #

This document collects short explanations of terminology one may encounter within the Lean community.

Whilst many of the entries below have precise technical definitions, preference is given to explanations of their conversational use, with additional references linked for further information.

A small number of interlinks found below represent entries which have not yet been added, and which therefore will not yet lead to entry definitions until they are completed. To request that an additional entry be added to this glossary, feel free to file an issue.

The entries on this page can be linked to via anchor links (e.g. https://leanprover-community.github.io/glossary.html#widget). For some entries, there are additional easier-to-type anchors which can be found just before the entry heading -- e.g. #heavy-rfl will lead to the "heavy rfl / heavy refl" entry. Authors of new glossary entries should consider adding these additional anchors if the entry title is long, contains backticks, or is otherwise hard to type.

attribute #

One or more tags or markings which may be applied to a Lean declaration, and which may affect either its behavior or the behavior of other Lean objects which interact with it. Attributes may be defined either within core Lean, within mathlib, or within any Lean code.

Applying an attribute is done either by prefixing a declaration command with @[name-of-attribute], or afterwards by using the attribute command, as in attribute [name-of-attribute] name-of-declaration.

The @[simp] attribute, for example, tags a declaration (typically a lemma, theorem or def) as being a simp lemma.

See also #

beta reduction #

A specific simplification operation in dependent type theory (and Lean's implementation of it) which may be performed as part of deciding whether two terms are definitionally equal.

More precisely, it is the process of simplifying an expression such as (λ x, t) a to t[a/x], where appearances of the variable x in t have been replaced with a.

See also #

big operators #

A locale within mathlib's algebra library, enabled via open_locale big_operators.

It defines notation for finite sums and products using the ∑ and ∏ characters.

binder #

Expressions such as (a : α), [a : α] or {a : α} for any identifier(s) a and type α which, as part of various Lean syntactical elements -- declarations, fun, quantifiers and others -- represent identifiers which will be bound within the body of the syntactic element or declaration.

Each type of binder has different implications for whether it will be bound implicitly (without being passed by a caller), explicitly, or via typeclass inference.

In some places, notably within a def, defining "simple" binders without surrounding brackets are allowed, such as a binder a (with no explicit type) for some identifier a.

bundled vs unbundled #

Given a mathematical object O with property P, bundling P refers to the creation of a Lean structure containing a proof of P as one of its fields in addition to those needed to structurally define O.

In contrast, an unbundled structure instead contains only the definition of O, and separately the creation of an is_P proposition which can be applied to terms of O.

As a concrete example, a group homomorphism can be viewed as a map φ: G → H between groups, along with a proof h : φ(a * b) = φ(a) * φ(b). A bundled group homomorphism would contain both φ and h as fields, whereas an unbundled one would instead consist only of φ, with a separate is_group_homomorphism declaration for the proof of h.

There are performance, stylistic or implementation-related reasons to prefer bundling or unbundling, as well as shades of grey which partially bundle some parts of a structure whilst leaving others unbundled. Type classes within mathlib are primarily semi-bundled, generally unbundling only the carrier type itself. Morphisms within mathlib are more often fully bundled, though traces of both approaches are present and discussed in the resource below.

See also #

cache #

Normally, a reference to a shared pre-built set of olean files built by mathlib's continuous integration each time a commit is pushed to its repository.

Its purpose is to alleviate the need for each mathlib user to build (or rebuild) the same Lean files locally when doing so can take a significant amount of time (multiple hours on a modest computer).

The cache is built within the aforementioned continuous integration, and normally users of mathlib retrieve its built files using leanproject.

calc mode #

A mode which consists of sequences of successive transformations of expressions involving a transitive relation such as =, < or others tagged with trans attributes. It is entered via the calc keyword.

See also #

carrier #

For a Lean structure which bundles a type T along with some additional mathematical structure represented as additional fields (e.g. Group), it is the type T of underlying elements.

For a Lean structure which bundles a set S along with some additional properties represented as additional fields (e.g. subgroup), it is the underlying set S.

class #

More fully, a typeclass (or type class).

A Lean structure whose instances can be retrieved via typeclass inference.

Distinct from the use of class in object oriented languages -- the word usage in functional programming languages comes from the typeclasses of Haskell.

code linter #

A code linter is a linter which attempts to find unintentional mistakes in Lean code within mathlib. Concretely, mathlib includes a collection of Lean programs which check for a variety of potential issues which may be introduced in new mathlib code. New linters are occasionally introduced to detect additional classes of mistakes. Mathlib's continuous integration ensures that any new such code passes the defined linters. Linting can be disabled for a particular piece of code by using the nolint attribute. Some linter failures that predate the CI code linters are indicated in an autogenerated nolint file; the length of this file should tend to zero over time as these errors are fixed.

conv mode #

A submode of tactic mode which facilitates navigating within assumptions or goals, in order to rewrite or simplify targeted portions of them. It is entered from tactic mode via the conv keyword.

See also #

core Lean #

As differentiated from mathlib or other community-authored Lean code, core Lean (or the "core library") refers to the portions of Lean which ship with the distribution of Lean itself.

Historically, even the mathlib project itself was a part of core Lean, and was split off into its own separately maintained project afterwards to facilitate development speed.

Some fundamental declarations remain part of core Lean even after the split. Occasionally additional lemmas and definitions are still removed or migrated from the core community Lean 3 repository and into mathlib, should they conflict with newly developed mathlib code.

The current portions of core Lean 3 can be found in the Lean 3 community repository, and for Lean 4 similarly.

declaration #

A single Lean runtime object within a Lean environment.

Or, ambiguously, any of a number of Lean commands which may define or declare such objects.

Examples of such commands are the def, theorem, constant or example commands (and in Lean 3, the lemma command), amongst others.

Further detail can be found in the Lean documentation.

dependent type theory #

A type theory in which one additionally may have types which depend on a parameter, such as the type of "days of a calendar month", whose specific values depend on the particular month, given that different months have different numbers of days. Further examples may be found in the resources below. Lean's implementation of dependent type theory is based on what is known as the Calculus of Constructions, allowing for its use both in complex mathematical reasoning as well as software verification.

See Also #

diamond #

The existence of multiple conflicting terms of a class found within the typeclass instance graph. A diamond is likely to cause issues during typeclass inference, which attempts to construct a single term of the class and may therefore be unable to do so. When unqualified, "diamond" refers most often to this undesirable case, where non-defeq terms in the diamond may cause errors, lead to goals which are not provable via refl, or potentially ones not provably equal at all. Within mathlib, diamonds are abundant because of its many hierarchies. Fixing or mitigating diamonds often involves refactoring the fields or instance priorities for the offending class. Diamonds which cross library boundaries -- such as ones in which part of the typeclass graph lives within mathlib and part within a library depending on mathlib which adds new instances or classes -- may be particularly hard to fix or avoid without modification.

See Also #

dot notation / generalized field notation / generalized projections #

The syntax sugar allowing notations such as ((foo a b c).bar x.y).baz.

Lean provides two interpretations of the syntax foo.bar: it can mean the declaration bar in the foo namespace, or it can be generalized field notation. We expand here on the latter.

Suppose foo has type C x1 ... xn, with C some constant and x1 ... xn arbitrary, and suppose that there is a declaration in the context named C.bar which takes an argument of type C x1 ... xn. Then foo.bar is sugar for C.bar foo. For calls of the form foo.bar _ ... _ with (implicit or explicit) arguments, Lean is smart enough to expand to C.bar _ ... foo _ ... _, so that everything typechecks. In these previous examples, foo can also be a more complicated expression such as the function application in (foo bar baz).quux.

equiv #

As distinct from mathematical equality, equiv allows for defining an equivalence or congruence of types. One important thing to note is that an equiv holds data instead of being merely a proof.

goal #

Within the context of interactively proving theorems in Lean, each targeted statement whose proof is in-progress.

Or more broadly, type theoretically, an individual type for which a term is to be exhibited. For propositions, exhibiting a term reduces equivalently to the aforementioned idea of proof.

golfing #

An attempt or effort to make a particular working piece of code as short as possible; within the context of Lean, often it is a proof's length which is golfed or optimized. Golfed proofs often make use of term mode, and in general prioritize terseness over readability. This obfuscation often is used intentionally or advantageously to signal to readers that a proof is mechanical or trivial.

heavy rfl / heavy refl #

A use of rfl (or refl) which performs slowly when it is evaluated by Lean. A heavy rfl occurs when rfl is asked to perform many steps of definitional reduction at once, resulting in a proof term which is small but which requires Lean to do a lot of computation to ensure that it type checks.

This Zulip discussion has a particularly slow example.

hierarchy #

A collection of successively more constrained typeclasses within a related area of mathematics. In mathlib, we have the algebraic hierarchy (semiring, ring, field, ...), the order hierarchy (preorder, partial_order, linear_order, ...), the topological hierarchy (t1_space, t2_space, normal_space, ...), the categorical hierarchy (preadditive, abelian, monoidal, ...) but also the scalar hierarchy (mul_action, distrib_mul_action, module, ...), the norm hierarchy, and intersections of previous ones like the order-algebraic hierarchy, the topologico-algebraic hierarchy, and others.

HoTT #

Homotopy Type Theory, a type theory distinguished by the inclusion of an additional univalence axiom, which makes precise the notion that two disparate implementations of a type which would be thought of as equivalent are also in a mathematical sense equal.

In Lean 2, there was native support within core Lean itself for a side-by-side homotopy type theory-based library.

Lean 3's kernel introduced singleton elimination, which is inconsistent with HoTT's univalence axiom. As a result, core Lean 3 no longer shipped a HoTT library. A partial port of the Lean 2 HoTT library for Lean 3 was made in an external project by Gabriel Ebner and fellow contributors.

As an illustrative concrete example, there are many definitions of the type of natural numbers which can be viewed as constructing equivalent mathematical objects. Core Lean has a Peano-esque implementation of them, and mathlib has an additional binary representation-based one which is shown to be equivalent. Given that Lean does not have the univalence axiom, these two types are equivalent, but they are not provably equal as types. A HoTT-based language (or library) would additionally call these types equal.

Icc, Ico, Ioc, Ioo, Ici, Ioi, Iic, Iio #

Shorthand notation used within mathlib for referring to one of 8 kinds of mathematical intervals. There are 8 types of intervals, depending on whether the interval is closed, open or runs to infinity at each end. Names have been made compact by calling each interval I + how it ends on the left + how it ends on the right. Iii, which in the naming convention would refer to an interval infinite at both ends, is not in use.

See also #

infoview #

Within the context of interactively editing Lean files, a window or interface which displays incremental goal state, diagnostics, errors and Lean widgets output.

instance #

One of two closely related concepts:

kernel #

In the context of Lean's implementation (and proof assistants more generally), the kernel is the central component which verifies the correctness of each proof.

Lean's kernel implementation is relatively small compared to the size of Lean's tactic implementation code[1^], or certainly compared to mathlib, allowing for greater confidence that proofs are free of mistakes without reliance on large or complex higher level constructs.

[1^]: this relative smallness, and ergo independent verifiability, is known as the de Bruijn criterion for a proof assistant's implementation

See also #

Lean Together #

An annual meeting for the Lean user community, its mathlib library users, as well as the broader theorem prover community.

Previous talks or presentations from the events have been shared on the community YouTube channel.

leanproject #

A higher-level supporting tool for working with projects, particularly mathlib, in Lean 3. It lives within the community mathlib-tools repository.

See also #

lint #

A linter is a small program that looks for hard-to-spot mistakes in code. mathlib defines both style linters and code linters. Mathlib is linted on every CI run when pull requests are submitted.

mathlib #

A large, community maintained collection of mathematics for Lean 3.

In addition to its breadth of mathematical objects and proofs, mathlib serves as the de facto standard (programmatic) library for Lean 3, containing additional functionality useful for non-mathematics related projects.

mathlib4 #

An experimental attempt to port portions of mathlib to Lean 4. mathlib4 is a manual, exploratory attempt to suss out points of difficulty, identify areas for redesign, and prepare for a more targeted effort which will occur as part of moving mathlib to Lean 4. Learnings from both the mathlib4 effort as well as from mathport often lead to backported changes to mathlib, to bring Lean 3 code into more future-compatible states.

mathport #

An in-development tool for automated or semi-automated translation of Lean 3 code into Lean 4 code. It consists of both fully automated generation of Lean 4 olean files from Lean 3 source files (binport), as well as best-effort source-to-source translation of Lean 3 to Lean 4 source code (synport). Learnings from both the mathport effort as well as from mathlib4 often lead to backported changes to mathlib, to bring Lean 3 code into more future-compatible states.

mode #

In the context of writing Lean code, a set of related syntactical elements or keywords which make a particular style of formal reasoning or construction of proof terms efficient. Lean, especially as restricted to mathlib, has a small number of such modes -- tactic mode, term mode, calc mode and conv-mode. A particular mode may make progress towards specific kinds of goals easier. A proof may often, however, mix various modes in the course of constructing a proof term.

module #

A single file containing Lean source code.

Not to be mistaken with a module within mathematics, i.e. the generalization of a vector space.

module docstring #

Module-level comment summarizing what's to be found in the file. We require that every file has one, but some old files still don't.

MWE #

Minimal Working Example, a way of making it easier to get help with a snippet of Lean code by reducing it to its essential parts, whilst being still runnable by others.

Further information can be found on the MWE page.

non-terminal simp #

An invocation of the simp tactic which is not the last tactic invoked on a particular subgoal, nor uses simp only to explicitly limit which simp lemmas it considers. Non-terminal simps are avoided because they are hard to maintain, given that their behavior or runtime will change as mathlib adds or modifies the set of simp lemmas over time.

See also #

The "non-terminal simp" section of the simp documentation

olean file #

A cached, compiled binary file produced by Lean as part of building a Lean module. olean files are tied to the specific Lean version which built them, and their names match the module they correspond to (so a file named foo.lean will have a corresponding foo.olean file). Building an olean file can be done manually (via e.g. lean --make foo.lean), but for collaborative projects like mathlib they are built via CI as a shared olean cache, and then can simply be retrieved by each mathlib user.

orange bar of hell #

Interactive editing of Lean in VSCode (or other editors) is typically fairly responsive.

Orange bars of hell, however, refer to occasional instances where orange bars in the sidebar alongside a Lean file do not disappear or update. Normally these bars indicate which parts of a file Lean is still evaluating, but if these bars persist, it indicates that progress isn't being made. Fixing these bars can often be done by closing any inactive editor tabs, followed by opening the VSCode Command Palette (ctrl-shift-p or cmd-shift-p) and running Lean: Restart. Another common reason this occurs is if Lean is having to (re)compile all imported mathlib files due to a mismatched set of cached olean files. In these cases, ensuring that the mathlib cache is properly downloaded via leanpkg configure && leanproject get-mathlib-cache should fix the issue.

propeq #

Propositional equality. Two terms a b : α are propositionally equal if we can prove a = b. This is weaker than definitional and syntactical equalities.

See also #

Equality, specifications and implementations, from the Xena Project

simp lemma #

A lemma which has been tagged with an attribute which enables it for use with the simp tactic.

Good simp lemmas guide the simp tactic towards reducing complex expressions into simpler ones, often to the point where the simp tactic is able to close many goals itself.

Further details can be found in the simp documentation of mathlib.

simp-normal form #

A convention within mathlib for expressing propositions with multiple equivalent forms in a single conventional one.

Examples and further detail can found on the simp page.

style linter #

A style linter is a linter which attempt to ensure uniform appearance or style of code within mathlib, without affecting its working behavior. Concretely, mathlib includes a short Python program which checks for instance that lines are less than 100 characters long, that every file has a module docstring, et cetera. Mathlib's continuous integration ensures that new code passes the defined linters. Allowed style linting exceptions are stored in a style exceptions file within the repository.

tactic mode #

A Lean mode characterized by its reliance on sequences of tactics which often facilitate proofs quite similar to paper-based reasoning, albeit often with the use of sophisticated tactics which automate tedious portions of a proof. There are various means to enter tactic mode. It may be entered using the by keyword from term mode, though in Lean 3 it is most often entered via a begin...end block whenever its body is made up of multiple commands. Other modes can also be interspersed within it, often to collaboratively produce an understandable, efficient, short or readable overall proof. Ultimately, the result of a tactic mode block is a term, assembled via the tactics within it.

See also #

term mode #

A Lean mode which assembles a single term through the use of functional subexpressions. In contrast to tactic mode, term mode proofs are often short in length, though potentially harder to read for humans. There are various means to enter term mode. The body of a declaration begins in term mode, or within tactic mode it is often entered using the exact tactic. Efficient term mode proofs often contribute to code golfing.

Commands such as have, suffices, and show can be used to write structured term mode proofs that are more legible than bare proof terms.

TPIL #

"Theorem Proving in Lean", a free online textbook by Jeremy Avigad, Leonardo de Moura, and Soonho Kong, "designed to teach you to develop and verify proofs in Lean". Starting with a simple introduction to the type theory used in Lean, the text proceeds to explain topics such as tactics, inductive types, and type classes.

type theory #

A formal system with two fundamental kinds of objects, terms and types. It also may refer to the field of study of such languages, as there are nuances between various additional properties or features a specific type theory may contain.

Lean's dependent type theory underlies its foundational system of mathematics in contrast to set theory. Foundationally in set theory, all objects are in a formal sense built out of a single kind of primitive object, sets. One reasons about whether constructed sets are or are not members of other sets. This membership notion may be asked about any two formal sets, which can allow statements to be formally meaningful even if they would be mathematically meaningless to a human -- such as asking whether the number 2 is a member of the number 37 (with numbers defined set theoretically). By contrast, in type theory, each term has an associated type, as indeed do statements or propositions themselves. One has for instance the type of natural numbers (in mathlib, ℕ), or the type of propositions (Prop), or the type of proofs of 2 + 2 = 4, or of functions from ℕ to Prop (ℕ → Prop), and can construct terms of these types -- perhaps the terms 2, 2 + 2 = 37, rfl, and λ n, true, respectively. Not every term is a type however, so in contrast to the aforementioned set theoretic difficulties, one may not construct statements asking whether 37 is of type 2 if doing so is not meaningful.

unicode abbreviation #

In the context of editing Lean files, an abbreviation is a way of entering a symbol not generally found on standard keyboard layouts using a descriptive shortcut.

For instance, the not-equal symbol "≠" can be entered using the sequence \neq.

The full list of abbreviations (and their replacements) can be found in the vscode-lean repository.

whnf #

A Lean expression is in weak head normal form, often abbreviated to whnf, if it meets a number of normalization criteria mentioned in the resources linked below. Informally, expressions in whnf have had their outermost parts evaluated, though inner subexpressions may not have been evaluated.

It also may refer to a command which reduces expressions to this form.

See also #

widget #

An extensible framework for defining interactive, componentized graphical elements via Lean code, which render in infoviews during interactive theorem proving.

A widget may also refer to an individual graphical element being rendered by the aforementioned framework.

Widgets provide a mechanism to show additional context or information which updates in the course of interactive theorem proving.

Support is present both in Lean 3 and Lean 4, albeit with different out-of-the-box functionality.

See also #