Library Reference
Rewrite
AlgebraicRewriting.CategoricalAlgebra.CSets.Migrate
— MethodApply migration to all literals in the constraint
AlgebraicRewriting.Rewrite.Constraints.BoolAnd
— TypeConjunction of multiple expressions
AlgebraicRewriting.Rewrite.Constraints.BoolConst
— TypeConstant, independent of context
AlgebraicRewriting.Rewrite.Constraints.BoolExpr
— TypeSomething that, in a context, can be evaluated to a bool
AlgebraicRewriting.Rewrite.Constraints.BoolNot
— TypeNegation of an expression
AlgebraicRewriting.Rewrite.Constraints.BoolOr
— TypeDisjunction of multiple expressions
AlgebraicRewriting.Rewrite.Constraints.CGraph
— Type"nothing" means something that will be determined via a quantifier Ints are explicit arguments provided when apply_constraint is called
AlgebraicRewriting.Rewrite.Constraints.Commutes
— TypeA commutative diagram with multiple parallel paths, asserted to either commute or to not commute
AlgebraicRewriting.Rewrite.Constraints.Constraint
— TypeA constraint graph and a BoolExpr (which refers to the constraint graph)
AlgebraicRewriting.Rewrite.Constraints.Quantifier
— TypeQuantified edge
e - which edge is filled in kind - Exists, Forall, or Exists! st - "such that", restrict the domain of quantification via a condition monic - restrict domain of quanitification to only monic matches
AlgebraicRewriting.Rewrite.Constraints.AppCond
— FunctionConstraint a constraint that asserts (or denies) the existence of a triangle commuting.
f₁
(1) <- (2) ∃₂↘ ↓ λ₃ (3)
AlgebraicRewriting.Rewrite.Constraints.LiftCond
— Method ∀₂
(1) → (3) ₁↓ ↗∃₃ ↓ λ₅ (2) → (4) ⁴
Test a map (3)→(4), given maps (1)->(2)->(4).
AlgebraicRewriting.Rewrite.Constraints.arity
— MethodNumber of variables in a constraint graph
AlgebraicRewriting.Rewrite.Constraints.check_expr
— MethodValidate a commutative diagram constraint makes sense
AlgebraicRewriting.Rewrite.Constraints.eval_boolexpr
— MethodCheck whether homs are equal by looping over domain.
AlgebraicRewriting.Rewrite.Constraints.get_ob
— MethodGet the C-Set associated with a vertex in a CGraph
AlgebraicRewriting.Rewrite.Constraints.merge_graphs
— MethodTake two CGraphs and merge them along their overlapping vertices and edges Returns an ACSetColimit
Catlab.Theories.:⊕
— MethodCombine two constraints disjunctively, sharing as much of the computation graph as possible.
Catlab.Theories.:⊗
— MethodCombine two constraints conjunctively, sharing as much of the computation graph as possible (i.e. pushout along the maximum common subgraph)
AlgebraicRewriting.Rewrite.Utils.Rule
— TypeRewrite rules which are (usually) encoded as spans. The L structure encodes a pattern to be matched. The R morphism encodes a replacement pattern to be substituted in. They are related to each other by an interface I with maps: L ⟵ I ⟶ R
A semantics (DPO, SPO, CoNeg, or SqPO) must be chosen.
Control the match-finding process by specifying whether the match is intended to be monic or not, as well as an optional application condition(s)
AlgebraicRewriting.Rewrite.Utils.can_match
— MethodReturns nothing if the match is acceptable for rewriting according to the rule, otherwise returns the reason why it should be rejected
AlgebraicRewriting.Rewrite.Utils.freevars
— MethodGet a list of AttrVar indices which are NOT bound by the I→R morphism
AlgebraicRewriting.Rewrite.Utils.get_expr_binding_map
— MethodDon't bind variables for things that are not ACSets
AlgebraicRewriting.Rewrite.Utils.get_expr_binding_map
— MethodGiven the match morphism and the result, construct a map X → X′ which binds any free variables introduced into the result.
L <- I -> R m ↓ ↓ ↓ res G <- • -> X ↓ X′
AlgebraicRewriting.Rewrite.Utils.get_match
— MethodGet one match (if any exist) otherwise return
AlgebraicRewriting.Rewrite.Utils.get_matches
— MethodGet list of possible matches based on the constraints of the rule
This function has the same behavior as the generic get_matches
, but it is more performant because we do not have to query all homomorphisms before finding a valid match, in case n=1.
AlgebraicRewriting.Rewrite.Utils.get_matches
— MethodIf not rewriting ACSets, we have to compute entire Hom(L,G).
AlgebraicRewriting.Rewrite.Utils.get_pmap
— MethodExtract the partial map (derived rule) from full output data
AlgebraicRewriting.Rewrite.Utils.get_rmap
— MethodExtract the map from the R to the result from the full output data
AlgebraicRewriting.Rewrite.Utils.rewrite
— Methodrewrite(r::Rule, G; kw...)
Perform a rewrite (automatically finding an arbitrary match) and return result.
AlgebraicRewriting.Rewrite.Utils.rewrite_match
— Methodrewrite_match(r::Rule, m; kw...)
Perform a rewrite (with a supplied match morphism) and return result.
AlgebraicRewriting.CategoricalAlgebra.CSets.var_eqs
— MethodFurther induced equations between AttrVars, given a specific match morphism
AlgebraicRewriting.Rewrite.Utils.check_match_var_eqs
— MethodIgnore for other categories
AlgebraicRewriting.Rewrite.Utils.check_match_var_eqs
— MethodA match may be invalid because two variables (which are to be assigned different values via the I -> R map) are identified (due to merging via the I->L map, which morally ought be monic but is not for AttrVars). We can check this before computing the pushout to make sure that we will not get an inconsistent result when trying to compute it. This requires executing the custom exprs of the rewrite rule, so we may wish to build in the ability to skip this step if that is computationally intensive.
AlgebraicRewriting.Rewrite.Utils.rewrite_match_maps
— Methodrewrite_match_maps(r::Rule{:DPO}, m)
Apply a DPO rewrite rule (given as a span, L<-I->R) to a ACSet using a match morphism m
which indicates where to apply the rewrite. l r L <- I -> R m ↓ ↓ ↓ G <- K -> H
This works for any type that implements pushout_complement
and pushout
AlgebraicRewriting.Rewrite.Utils.rewrite_match_maps
— Methodrewrite_match_maps(r::Rule{:CoNeg}, m)
Apply a CoNegation rewrite rule (given as a span, L↩I->R) to a ACSet using a monic match morphism m
which indicates where to apply the rewrite. l r L <- I -> R m ↓ ↓ ↓ G <- K -> H where K = ~L ∨ I
This works for any type that implements bi-Heyting logic operators ~ and ∨.
This is described here. Essentially, it is partway between DPO and SPO. Suppose the rule tries to delete two things, one of which satisfies the dangling condition, the other violates it. While DPO would fail to apply at all, and SPO would delete both things (cascading the deletion for the latter), co-negation rewriting would simply delete the item which can be deleted without cascading and ignore the other element.
It includes a quote which indicates that this method should work even when the match morphism isn't monic, if it satisfies the identification condition. Supporting this is not yet implemented.
Match morphisms which bind attribute variables are not monic, hence we this form of rewriting doesn't support VarACSets. Intuitively, it feels like this restriction could be relaxed.
AlgebraicRewriting.Rewrite.SPO.partial_pushout
— MethodC ← Ag ↪ A ↩ Af → B
A ↩ f∇g → Bgf ↪ B ↓ ⌜ ↓ C ↩ Cfg -> D
Implementation of Construction 6 in Löwe's "Algebraic approach to SPO graph transformation"
AlgebraicRewriting.Rewrite.SqPO.final_pullback_complement
— MethodSee Theorem 2 of 'Concurrency Theorems for Non-linear Rewriting Theories' f B <–- A m ↓ ↓ n C <– D g
AlgebraicRewriting.Rewrite.Utils.rewrite_match_maps
— Methodrewrite_match_maps(r::Rule{:SqPO},m; pres::Union{Nothing, Presentation}=nothing)
Sesqui-pushout is just like DPO, except we use a final pullback complement instead of a pushout complement.
r.L r.R
L <-⌞K -> R m ↓ ↓k ↓ r I <- • ->⌜O i o
AlgebraicRewriting.Rewrite.PBPO.PBPORule
— Type l r
L ⟵ K ⟶ R tl ↓ ↓ tk <== tl, tk must be monic L' ⟵ K'
It is assumed we never want the typing/adherence match to be monic, but we can optionally restrict the match L → G to be monic.
We can attach application conditions to both the match morphism as well as the adherence morphism. Until morphism search under constraints becomes efficient, it's sometimes needed to just directly state the adherence morphism as a function of the match morphism.
AlgebraicRewriting.Rewrite.PBPO.canon
— MethodTake a PBPO rule and put into normal form, i.e. where the lower square forms a pullback
See Prop 2.4 of "The PBPO graph transformation approach"
AlgebraicRewriting.Rewrite.PBPO.partial_abstract
— MethodThis construction addresses the following problem: ideally when we 'abstract' an ACSet from X to A->X, maps into X, say B->X, can be canonically pulled back to maps B->A which commute. However, A won't do here, because there may not even exist any maps B->A. If B has concrete attributes, then those cannot be sent to an AttrVar in A. Furthermore, if B has multiple 'references' to an AttrVar (two different edges, each with AttrVar(1), sent to two different edges with the same atttribute value in X), then there is no longer a canonical place to send AttrVar(1) to in A, as there is a distinct AttrVar for every single part+attr in X. So we need a construction which does two things to A->X, starting with a map B->X. 1.) replaces exactly the variables we need with concrete values in order to allow a map B->A, 2.) quotients variables in A so that there is exactly one choice for where to send attrvars in B such that the triangle commutes.
Starting with a map L -> G (where G has no AttrVars), we want the analogous map into a "partially abstracted" version of G that has concrete attributes replaced with AttrVars EXCEPT for those attributes which are mapped to by concrete attributes of L. Likewise, multiple occurences of the same variable in L correspond to AttrVars which should be merged in the partially-abstracted G.
For example, for a schema with a single Ob and Attr (where all combinatorial maps are just {1↦1, 2↦2}):
L = [AttrVar(1), :foo]
G = [:bar, :foo, :baz]
abs(G) = [AttrVar(1), AttrVar(2), AttrVar(3)]
expected result: [AttrVar(1), :foo, AttrVar(2)]
L -> Partial_abs(G) ↓ ↑ G <- abs(G)
This function computes the top arrow of this diagram starting with the left arrow. The bottom arrow is computed by abstract_attributes
and the right arrow by sub_vars
. Furthermore, a map from Partial_abs(G) to G is provided.
This is the factorization system arising from a coreflective subcategory.
(see https://ncatlab.org/nlab/show/reflective+factorization+system and https://blog.algebraicjulia.org/post/2023/06/varacsets/)
AlgebraicRewriting.Rewrite.Utils.get_expr_binding_map
— MethodUse exprs and k_exprs to fill in variables introduced by applying the rw rule.
AlgebraicRewriting.Rewrite.Utils.get_matches
— MethodPBPO matches consist of two morphisms. First, a match m: L → G, and secondly a typing G → L′. With attributes, it is not so simple because G has concrete values for attributes and L′ may have variables. Therefore, we actually change the typing to map out of A, an abstracted version of G (with its attributes replaced by variables). So we lift matches L->G to matches L->A, then search α∈Hom(A,L′).
In general, we want α to be uniquely determined by m, so by default α_unique
is set to true.
m
L⌟ ⟶ G || ↓ α L ⟶ L′ tl
m
L ⟶ G tl ↓ ↘a ↑ (abs = partial abstraction. Note a
is Labs
in the code.) L′⟵ A α
The "strong match" condition we enforce is that: tl⁻¹(α(A)) = a⁻¹(A). This means we can deduce precisely what m is by looking at α.
AlgebraicRewriting.Rewrite.Utils.rewrite_match_maps
— Method r
+ K ----> R
+gₗ u ↓ gᵣ ⌜ ↓ w
Gₗ <–– Gk ––> Gᵣ α ↓ ⌞ ↓ u' L′ <– K′ tₗ
For the adherence morphism α to be valid, it must satisfy a condition with m, tₗ. This is checked for matches provided by get_matches, so by default we do not check it.
L <–⌞• m ↓ ↓ G ⟵ Gk
See Lemma 7.2 of "TERMINATION OF GRAPH TRANSFORMATION SYSTEMS USING WEIGHTED SUBGRAPH COUNTING" by Overbeek and Endrullis (2023)
AlgebraicRewriting.Rewrite.Migration.pres_hash
— MethodWant a filename that is stable to multiple Julia sessions but changes when the schema changes. This minimizes the need to clear the cache.
Schedules
AlgebraicRewriting.Schedules.Poly
— ModuleMealy machines (augmented with monadic output) are a user-friendly format for specifying a behavior tree. Behavior trees in general are not finitely expressible, but we focus on trees which can be lazily generated by functions.
Although it is conceptually simple to think of a single set of "input doors" out "output doors" to enter/leave the Mealy machine, such that a Mealy machine has type A → B, we use Σᵢ Aᵢ → Σⱼ Bⱼ, where Aᵢ and Bⱼ are Julia types. This allows us to represent a Mealy machine with (Int + String)-many input doors, for example.
AlgebraicRewriting.Schedules.Poly.List
— Constant"The list monad returns the set of packages labeled with a natural number N, each of which has N-many slots."
AlgebraicRewriting.Schedules.Poly.Maybe
— Constant"The Maybe monad, y+1, consists of two packages, one with one slot and the other with no slots."
AlgebraicRewriting.Schedules.Poly.BTree
— TypeLazily grown behavior tree induced by a Mealy machine. The future behavior is dictated by the inputs seen thus far (i.e. a vector of WireVals).
Each vertex is identified by a sequence of inputs and has a state of the Mealy machine associated with it.
Each nonempty sequence of inputs has a MealyRes associated with it.
AlgebraicRewriting.Schedules.Poly.BTree
— MethodGrow a tree and return the result.
AlgebraicRewriting.Schedules.Poly.Mealy
— TypeA function that maintains a state (initially s0) and has monadic output for some polynomial monad t.
The function f must be of type S × WireVal → S × (t ◁ WireVal)
(i.e. S × Wireval → MealyRes)
AlgebraicRewriting.Schedules.Poly.MealyRes
— TypeOutput of a Mealy machine
newS - The new state of the Mealy machine mval - outputs along with their monadic values (e.g. probability weights) msg - A message reporting something about the computation
AlgebraicRewriting.Schedules.Poly.PMonad
— TypeQuotes in this docstring and others taken from David Spivak: https://topos.site/blog/2023/09/powers-of-polynomial-monads/#exponentiating-monads
"A polynomial monad is a polynomial functor t with coherent maps η: y → t and μ: t ◁ t → t.
Polynomial monads can be thought of as offering compositional (possibly labeled) packages with some number of slots. The compositionality of this packaging says that (via the monad unit) we know how to package up a given element of any set, and that (via the monad multiplication) we can take a package of packages and simplify it to a single package."
We consider monads t of the form: t = Σ_{i ∈ t(1)} y^{t[i]}
I is the type of labels, e.g. probability weights.
AlgebraicRewriting.Schedules.Poly.Simulator
— TypeEquips a wiring diagram description of a simulator with mutable data structures (now behavior trees for each box, but possibly incremental homomorphism caches in the future). Requires that all the boxes of the WiringDiagram be convertable to BTrees.
AlgebraicRewriting.Schedules.Poly.TrajStep
— TypeA trajectory step is a box being fed a particular value
This only makes sense with reference to a WiringDiagram which box
refers to.
AlgebraicRewriting.Schedules.Poly.WireVal
— TypeFor an in/output, Σᵢ Aᵢ, provide wire index + value on wire
AlgebraicRewriting.Schedules.Poly.add_edge
— MethodAppend without mutating
AlgebraicRewriting.Schedules.Poly.apply_schedule
— MethodIn theory applying a schedule should result in a list of ACSets associated with out ports and monad labels (e.g. probabilities), and if one were to want to recover the trajectory of the output one would have to use a Writer monad of some sort. For simplicity, the application of a schedule will simply return the trajectories themselves (monadic multiplication could in principle condense this output to the pure output).
AlgebraicRewriting.Schedules.Poly.apply_traj_step
— MethodA particular trajectory enters a box. Out of the box comes a list of trajectories.
AlgebraicRewriting.Schedules.Poly.curr_state
— MethodCurrent state of the world
AlgebraicRewriting.Schedules.Poly.currwire
— MethodGet the wire which the traj is currently on
AlgebraicRewriting.Schedules.Poly.joindist
— Method"The lotteries monad returns the set of packages labeled with a lottery (a natural number N and a probability distribution on it) and again containing N-many slots."
AlgebraicRewriting.Schedules.Poly.joinstr
— MethodWriter monad
AlgebraicRewriting.CategoricalAlgebra.CSets.Migrate
— MethodMap a functor over the data of a schedule
AlgebraicRewriting.Schedules.Wiring.AgentBox
— TypeType for primitive boxes used in a schedule. These are the generating morphisms of a traced monoidal category, with objects being lists of ACSets.
ACSets.DenseACSets.sparsify
— MethodMap sparisfication over the data of a schedule
AlgebraicRewriting.Schedules.Wiring.merge_wires
— FunctionThe comonoid structure - merging multiple wires into one. This is unproblematic because the world state only ever exists on one wire at a given time.
AlgebraicRewriting.Schedules.Wiring.mk_sched
— MethodMake a wiring diagram with ob/hom generators using @program macro
TODO double check that this does not introduce any wire splitting.
AlgebraicRewriting.Schedules.Wiring.singleton
— MethodMake a wiring diagram around a box
AlgebraicRewriting.Schedules.Wiring.str_hom
— MethodVisualize the data of a CSet homomorphism
AlgebraicRewriting.Schedules.Wiring.wire_vals
— Method1 = inwire, 2 = outwire
AlgebraicRewriting.Schedules.Eval
— ModuleSpecialized code for handling rewriting of ACSets with the identity monad
AlgebraicRewriting.Schedules.Eval.interpret!
— Methodinterpret a wiring diagram, with each box updating its state in place
AlgebraicRewriting.Schedules.Eval.interpret
— MethodInterpret a wiring diagram, recording the trajectory taken
AlgebraicRewriting.Schedules.Basic.Initialize
— TypeA box that spits out a constant ACSet with an empty agent above it. Possibly, it does not take any inputs, so it can act as a comonoid counit.
AlgebraicRewriting.Schedules.Basic.Strengthen
— TypeAdds to both agent and the state of the world via a pushout.
Agent₁ → Agent₂
+ ↓ ⇣
+ World₁ -->⌜World₂
AlgebraicRewriting.Schedules.Basic.Weaken
— TypeChange the agent to a subobject of the current agent without changing the world
AlgebraicRewriting.Schedules.Conditionals.Conditional
— TypeA primitive box in a NestedDWD which does not change the state but redirects it out of one of n wires.
It contains a function (A->X) -> ℝⁿ. This optionally depends on the internal state. This weights probability for n outports, conditional on the status of an ACSet. If the function just depends on X rather than the whole morphism, withagent
is false
. If the function does not depend on the internal state (assumed to be true iff initial state is nothing
), then withstate
is false
.
The state and update function are by default trivial.
AlgebraicRewriting.Schedules.Conditionals.const_cond
— MethodCreate a branching point with fixed probabilities for each branch
AlgebraicRewriting.Schedules.Conditionals.for_schedule
— MethodPerform a 1-1 schedule n
times
AlgebraicRewriting.Schedules.Conditionals.if_cond
— MethodEnter the 1st branch iff the world state evaluates to true
AlgebraicRewriting.Schedules.Conditionals.uniform
— MethodA uniform chance of leaving each of n branches
AlgebraicRewriting.Schedules.Conditionals.while_schedule
— MethodPerform a 1-1 schedule until a condition is met
AlgebraicRewriting.Schedules.RuleApps.RuleApp
— TypeHas the semantics of applying the rule to some match that is found (no guarantees on which one, which should be controlled by application conditions). If rewrite occurs, exit mode 1, else exit mode 2.
The agent is related to the L and R patterns of the rule. This can be done via a Span, or implicitly as a homomorphism into "I" of the rewrite rule, and alternatively just from the shape of the agent alone (if it is identical to I, take the id map, otherwise take the unique morphism into I).
AlgebraicRewriting.Schedules.RuleApps.has_match
— MethodA box that takes the first output iff there is a match from a rule into the current state
AlgebraicRewriting.Schedules.RuleApps.loop_rule
— MethodFeed the "rewrite applied" output back into the input of the rule application
AlgebraicRewriting.Schedules.Queries.Query
— TypeHas an A input/output and a B input/output (by default, the B input can be changed to some other type if needed).
A R ---------↖
+↓ ↓ []
⌜–––-⌝ [] | Query | [agent subroutine] ⌞–––-⌟ [] ↓ ↓ ↓ [] A B ∅ [] ↘–––––-↗ Performs one action per element of Hom(B,X), optionally with some constraints. (i.e. sends you out along the B wire with agent Bₙ->X).
After you have done this for all Bₙ, then you exit the A port (you need to update the A->X map, and, if at any point the agent was deleted, then you exit a third door typed by 0).
A constraint optionally will be applied to (1) the A->W<-B cospan of old agent and purported new agent. (the new agent is the first argument to the constraint)
CategoricalAlgebra
AlgebraicRewriting.CategoricalAlgebra.FinSets.id_condition
— MethodCheck identification condition for pushout complement of finite sets.
The identification condition says that the functions do not map (1) both a deleted item and a preserved item in L to the same item in G or (2) two distinct deleted items to the same item. It is trivially satisfied for injective functions.
Returns pair of iterators of
(1) a nondeleted item that maps to a deleted item in G (2) a pair of distinct items in L that are deleted yet mapped to the same item in G.
AlgebraicRewriting.CategoricalAlgebra.FinSets.pushout_complement
— MethodCompute a pushout complement of finite sets, if possible.
Given functions $l: I → L$ and $m: L → G$ to form a pushout square
l
L ← I m ↓ ↓k G ← K g
define the set $K := G / m(L / l(I))$ and take $g: K ↪ G$ to be the inclusion. Then the map $k: I → K$ is determined by the map $l⋅m: I → G$ from the requirement that the square commutes.
Pushout complements exist only if the identification condition is satisfied. An error will be raised if the pushout complement cannot be constructed. To check this in advance, use can_pushout_complement
.
AlgebraicRewriting.CategoricalAlgebra.CSets.Migrate
— TypeTODO: check if functorial
AlgebraicRewriting.CategoricalAlgebra.CSets.cascade_subobj
— MethodRecursively delete anything, e.g. deleting a vertex deletes its edge
AlgebraicRewriting.CategoricalAlgebra.CSets.check_pb
— MethodY i↘ f_ X → • g_ ↓ ⌟ ↓ f • → • g
Check whether (X, f,g) is the pullback of (f,g), up to isomorphism (i.e. the pullback of f and g produces (Y,π₁,π₂), where Y is isomorphic to X and i⋅f_ = π₁ & i⋅g_ = π₂.
AlgebraicRewriting.CategoricalAlgebra.CSets.fibers
— MethodEvery morphism induces a partition of the parts of the domain. This function finds every nontrivial partition (size greater than one element) for the objects of the schema.
AlgebraicRewriting.CategoricalAlgebra.CSets.gluing_conditions
— MethodCheck both id condition and dangling condition
AlgebraicRewriting.CategoricalAlgebra.CSets.invert_hom
— MethodInvert a morphism which may not be monic nor epic. When the morphism is not monic, an arbitrary element of the preimage is mapped to. When it is not epic, a completely arbitrary element is mapped to.
AlgebraicRewriting.CategoricalAlgebra.CSets.invert_iso
— FunctionInvert some (presumed iso) components of an ACSetTransformation (given by s)
AlgebraicRewriting.CategoricalAlgebra.CSets.sub_vars
— FunctionGiven a value for each variable, create a morphism X → X′ which applies the substitution. We do this via pushout.
O –> X where C has AttrVars for merge
equivalence classes ↓ and O has only AttrVars (sent to concrete values or eq classes C in the map to C.
subs
and merge
are dictionaries keyed by attrtype names
subs
values are int-keyed dictionaries indicating binding, e.g. ; subs = (Weight = Dict(1 => 3.20, 5 => 2.32), ...)
merge
values are vectors of vectors indicating equivalence classes, e.g. ; merge = (Weight = [[2,3], [4,6]], ...)
AlgebraicRewriting.CategoricalAlgebra.CSets.var_eqs
— MethodFurther induced equations between AttrVars, given a specific match morphism
AlgebraicRewriting.CategoricalAlgebra.CSets.var_pullback
— MethodTake an ACSet pullback combinatorially and freely add variables for all attribute subparts.
This relies on implementation details of abstract
.
AlgebraicRewriting.CategoricalAlgebra.FinSets.can_pushout_complement
— MethodCan a pushout complement be constructed for a composable pair?
Even in nice categories, this is not generally possible.
AlgebraicRewriting.CategoricalAlgebra.FinSets.pushout_complement
— MethodPushout complement: extend composable pair to a pushout square.
Pushout complements are the essential ingredient for double pushout (DPO) rewriting.
AlgebraicRewriting.CategoricalAlgebra.FinSets.pushout_complement
— Methodpushout_complement(f::SliceHom, g::SliceHom)
Compute a pushout complement in a slice category by using the pushout complement in the underlying category.
f
B <– A –-⌝ | ↘ ↙ | g| X | f′ ↓ ↗ ↖ cx | D <–- C <– g′
AlgebraicRewriting.CategoricalAlgebra.FinSets.pushout_complement
— MethodCompute pushout complement of attributed C-sets, if possible.
The pushout complement is constructed pointwise from pushout complements of finite sets. If any of the pointwise identification conditions fail (in FinSet), this method will raise an error. If the dangling condition fails, the resulting C-set will be only partially defined. To check all these conditions in advance, use the function can_pushout_complement
.
In the absence of AttrVars, K is a subobject of G. But we want to be able to change the value of attributes. So any variables in I are not concretized by the I->K map. However, AttrVars may be merged together if m: L -> G
merges parts together.
Catlab.CategoricalAlgebra.HomSearch.homomorphisms
— MethodThis could be made more efficient as a constraint during homomorphism finding.
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.StructuredMultiCospanHom
— TypeA component-wise map between two cospans. The first component given is the apex map, with the following maps being the legs.
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.openrule
— TypeA span of StructuredMulticospanHoms, interpreted as a DPO rewrite rule
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.composeH_
— MethodcomposeH_(r₁, r₂)
compose two rewrite rules horizontally (via pushouts) as shown below: L₁₋₍ₙ₋₁₎-> L <- Lₙ X₁ -> X <- X₂₋ₘ L₁₋₍ₙ₋₁₎ -> L +Lₙ X <- X₂₋ₘ ↑ λ ↑ ↑ ↑ ↑ χ ↑ ↑ ↑ ↑ I₁₋₍ₙ₋₁₎-> I <- Iₙ ∘h Y₁ -> Y <- Y₂₋ₘ = I₁₋₍ₙ₋₁₎ -> I +Iₙ Y <- Y₂₋ₘ ↓ ρ ↓ ↓ ↓ ↓ ζ ↓ ↓ ↓ ↓ R₁₋₍ₙ₋₁₎-> R <- Rₙ Z₁ -> Z <- Z₂₋ₘ R₁₋₍ₙ₋₁₎ -> R +Rₙ Z <- Z₂₋ₘ
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.composeH_
— MethodCospan composition given by pushout
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.composeV_
— MethodFinset span composition given by pullback
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.composeV_
— MethodcomposeV_(r₁, r₂)
compose two rewrite rules vertically with pullbacks, as shown below: L₁₋ₙ -> L ↑ ↑ I₁₋ₙ -> I ↓ ↓ L₁₋ₙ -> L R₁₋ₙ -> R ↑ ↑ ∘v = I₁₋ₙ ×ᵣ₁₋ₙ Θ₁₋ₙ -> I ×ᵣ Θ Λ₁₋ₙ -> Λ ↓ ↓ ↑ ↑ Ω₁₋ₙ -> Ω Θ₁₋ₙ -> Θ ↓ ↓ Ω₁₋ₙ -> Ω
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.id2H_
— MethodPass dummy value in because a span of invertible FinFunctions does not retain L type
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.idV_
— MethodVertical arrows are spans of invertible finfunctions
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.open_pushout_complement
— MethodInitial data: 4 structured cospans + 3 cospan morphisms: μ, λ, ρ g G₁₋ₙ –> G ↑ l ↑ μ L₁₋ₙ –> L ↑ i ↑ λ I₁₋ₙ –> I ↓ r ↓ ρ R₁₋ₙ –> R
Computed data: 2 new structured cospans + 4 cospan morphisms: γ, η, ik, rh G₁₋ₙ G ↑ k ↑ γ ik I₁₋ₙ -> K₁₋ₙ –> K <– I ↓ h ↓ η rh R₁₋ₙ -> H₁₋ₙ –> H <– R In the context of the legs of a multicospan, the indices 1-n refer to the n legs of the cospan. In the context of a map of multicospans, there are 1-(n+1) maps, with the first one designating the map of the apexes. Hence it can make sense to have the elements: zip(legs, maps[2:end]) = [(legᵢ, mapᵢ), ...]
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.open_rewrite
— MethodApply a rewrite rule to a structured multicospan, where a matching cospan homomorphism is found automatically. If multiple matches are found, a particular one can be selected using m_index
. Returns nothing
if none are found.
AlgebraicRewriting.CategoricalAlgebra.StructuredCospans.open_rewrite_match
— MethodExtract the rewritten structured cospan from the induced rewrite rule
Catlab.CategoricalAlgebra.HomSearch.homomorphisms
— MethodFind homomorphisms between structured cospans. These are constrained to be iso on the legs of the cospans. Solving this w/ homomorphism finding requires a dynamic acset, and the current hack will be replaced once those are available.
A homomorphism backend that uses SAT/SMT would also make this viable to do without hacking.
AlgebraicRewriting.CategoricalAlgebra.PartialMap.check_eqs
— MethodConfirm a C-Set satisfies its equational axioms
AlgebraicRewriting.CategoricalAlgebra.PartialMap.eval_path
— MethodTake a GATExpr (an id morphism, a generator, or a composite) and evaluate, starting at a particular point.
AlgebraicRewriting.CategoricalAlgebra.PartialMap.partial_map_classifier_eta
— MethodThe natural injection from X ⟶ T(X) When evaluated on the terminal object, this gives the subobject classfier.
AlgebraicRewriting.CategoricalAlgebra.PartialMap.partial_map_classifier_universal_property
— MethodA partial function is defined by the following span: m f A ↩ X → B
We compute ϕ(m,f): A ⟶ T(B) such that the following is a pullback square: f X ⟶ B m ↓ ↓ η(B) A ⟶ T(B) ϕ
Essentially, ϕ sends elements of A to the 'real' values in T(B) when A is in the subobject picked out by X. When A is 'deleted', it picks out the right element of the additional data added by T(B).
AlgebraicRewriting.CategoricalAlgebra.PartialMap.partial_map_functor_hom
— MethodBecause the functorial embedding of objects keeps a copy of the original data, what to do with morphisms is just carry them along. Because our implementation adds all of the additional stuff afterwards, index-wise, we can use literally the same data for a morphism lifted from X⟶Y to T(X)⟶T(Y).
However, we still need to map the extra stuff in T(X) to the proper extra stuff in T(Y).
AlgebraicRewriting.CategoricalAlgebra.PartialMap.partial_map_functor_ob
— MethodA functor T, playing the role of Maybe in Set, but generalized to C-Sets.
When called on the terminal object, this produces the subobject classifier: See Mulry "Partial map classifiers and cartesian closed categories" (1994)
This function specifies what T does on objects. The key properties:
- for all X ∈ Ob(C), η(X):X⟶T(X) is monic. m f ϕ(m,f)
- for each span A ↩ X → B, there exists a unique morphism A ⟶ T(B) such that (m,f) is the pullback of ϕ(m,f),η(B))
Not only do we add an extra element to each component of the C-Set, but we need to consider the possibility that a component (with n outgoing morphisms) has any combination of the targets of those morphisms deleted (like the subobject classifier, there are different ways for something to be deleted).
For example, in Graph, an edge can be deleted that goes between any two vertices of the graph. We can't map all deleted edges to the same point in T(E) (if we're going to satisfy that desired property #2), so we need an extra edge in T(E) for every possibility (from V1 to V2, from V1 to V3, ..., from [Deleted] to V1, ..., from V2 to [Deleted], ... from [Deleted] to [Deleted]), where [Deleted] is our name for the extra element added to T(V).
[src] [tgt]
Thus, T(E) ≅ |E| + (|V|+1) × (|V|+1).
In general, T(X) ≅ |X| + ∏ₕ(|T(codom(h))|) for each outgoing morphism h::X⟶Y
- the |X| corresponds to the 'real' elements of X
- the second term corresponds to the possible ways an X can be deleted.
- This recursive formula means we require the schema of the C-set to be acyclic otherwise the size is infinite (assumes schema is free).
AlgebraicRewriting.CategoricalAlgebra.PartialMap.topo_obs
— MethodGet topological sort of objects of a schema. Fail if cyclic.