Comparisons with Other Packages

Bindings for Tetrad exist in https://github.com/cmu-phil/r-tetrad, but not in the form of a formal R package. The authors instead recommend using the Python bindings from R instead.↩︎

caugi exposes a general PDAG class that can represent any CPDAG, and is_cpdag() validates the structure post-hoc, but there is no dedicated CPDAG class and CPDAG invariants are not enforced at construction (e.g. that the graph is the completion of a Markov equivalence class).↩︎

caugi accepts class = "AG" (ancestral graph) in the constructor and exports is_mag() to validate maximality post-hoc, but there is no dedicated MAG class—MAGs are representable as AGs and checked rather than enforced.↩︎

igraph has no dedicated DAG class. is_dag() tests whether a directed graph is acyclic, but the package’s single igraph object class does not enforce DAG invariants. A directed graph that happens to be acyclic is representable, but the type is not distinguished at the object level.↩︎

igraph has no CPDAG, MPDAG, MAG, PAG, ADMG, or SWIG class and no associated type-level operations (confirmed absent from R source and NAMESPACE).↩︎

Undirected graphs are representable via is_directed = FALSE (e.g. make_empty_graph(directed = FALSE)), but there is no dedicated UG class enforcing undirectedness as a type invariant distinct from the general igraph object.↩︎

igraph supports per-edge attributes but has no dedicated mixed-edge class with typed semantics. Representing a true mixed graph requires manual attribute manipulation with no type enforcement.↩︎

graph has no dedicated DAG class. graphNEL and graphAM can hold directed graphs (via edgemode = "directed") that happen to be acyclic, but the package enforces no acyclicity invariants and provides no DAG-specific construction or validation.↩︎

No dedicated UG class. Undirected graphs are represented by setting edgemode = "undirected" on a general graphNEL / graphAM object; the package stores reciprocal edges internally but does not expose a named UG class.↩︎

MultiGraph holds a shared node set with multiple edge sets, each independently directed or undirected. It does not support a single edge that is both directed and bidirected, and is not designed around causal mixed-edge semantics (e.g. MAG/PAG/ADMG).↩︎

gRbase provides a dag() constructor and is_dag() type-check, with the underlying representation being an igraph object or adjacency matrix. There is no dedicated DAG class enforcing acyclicity invariants at the object level.↩︎

Same reasoning as [^grbase-dag]: ug() constructs an undirected graph and is_ug() tests for it, but the result is a plain igraph or matrix.↩︎

pcalg produces DAGs (via randomDAG() / randDAG() returning graphNEL objects, or the ParDAG / GaussParDAG class), and isValidGraph(amat, type = "dag") validates them. The dominant user-facing representation is an untyped graphNEL rather than a pcalg-owned DAG class.↩︎

dag2cpdag() produces a CPDAG and isValidGraph(amat, type = "cpdag") validates one; pcAlgo objects carry a graph with attribute "amat.cpdag". There is no dedicated exported CPDAG class.↩︎

addBgKnowledge() augments a CPDAG with background knowledge and re-applies Meek’s rules to produce an MPDAG. As with CPDAG, no dedicated exported MPDAG class exists; the object is a bare adjacency matrix with convention-level typing only.↩︎

MAG-type adjacency matrices (0/2/3 coding from Richardson & Spirtes 2002) are used throughout pcalg (e.g. dsepAM(), backdoor(), adjustment(amat.type = "mag"), pcalg2dagitty()), but there is no exported MAG class.↩︎

fciAlgo objects hold a PAG as an adjacency matrix in the @amat slot. fci() / rfci() / fciPlus() / dag2pag() all return fciAlgo. The class is exported but is a result container for FCI-family algorithms, not a freely constructable PAG class.↩︎

dagitty recognises only "dag", "pdag", "mag", "pag" via graphType(). equivalenceClass() calls dagToCpdag internally and is documented as producing a CPDAG, but the result is typed "pdag"—no distinct CPDAG class or invariant enforcement. MPDAG is likewise representable as "pdag" without Meek-completion logic. "ug", "admg", "swig" are absent from both the R API and the JS engine.↩︎

cpdag() computes the CPDAG of a DAG and valid.cpdag() validates structure, but bnlearn uses a single generic bn class for all graph types; no dedicated CPDAG class exists and CPDAG invariants are not enforced at the object level.↩︎

skeleton() produces a fully undirected graph and valid.ug() validates whether a bn object is completely undirected, but the result is still a plain bn object—no dedicated UG class with enforced undirected invariants exists.↩︎

A bn object can hold a mix of directed and undirected arcs (representing a PDAG), making mixed graphs implicitly representable, but there is no explicit untyped mixed-edge graph class distinct from the PDAG interpretation.↩︎

ggm represents DAGs as plain integer adjacency matrices (entries 0, 1, 10, 100) built with DAG(). There is no S3/S4 class enforcing DAG invariants; isAcyclic() must be called separately.↩︎

essentialGraph() converts a DAG to its essential graph (= CPDAG), returning an adjacency matrix. There is no dedicated CPDAG class; the result is an untagged matrix.↩︎

MAG() derives a maximal ancestral graph after marginalisation/conditioning, and isAG() validates ancestral-graph conditions. The representation is still an untagged adjacency matrix.↩︎

isADMG() checks whether an adjacency matrix satisfies ADMG conditions. No dedicated ADMG constructor exists; the matrix encoding is the same general integer format used for all mixed graphs.↩︎

Undirected graphs are built with UG(), which produces an adjacency matrix with edge-type value 10. No distinct UG class enforces undirected invariants.↩︎

makeMG() combines directed, undirected, and bidirected components into one adjacency matrix; grMAT() converts graphNEL/igraph inputs to this format. The representation is a general integer adjacency matrix with no commitment to a specific typed mixed-graph semantics.↩︎

MixedGraphs has no dedicated DAG class. is_DAG() checks that a mixedgraph carries only directed edges and is acyclic, but the constructor does not enforce DAG invariants. The mixedgraph object is an untyped container.↩︎

is_ADMG() tests whether a mixedgraph contains only directed and bidirected edges and is acyclic, but there is no dedicated ADMG class. ADMG-typed objects from the companion ADMGs package can be converted to/from mixedgraph via convert(), but MixedGraphs itself does not expose an ADMG class.↩︎

NetworkX’s DiGraph class stores directed edges and supports DAG-related algorithms (is_directed_acyclic_graph, topological_sort, ancestors, descendants), but the class itself does not enforce acyclicity as an invariant—a DiGraph can hold cycles. There is no dedicated DAG subclass.↩︎

NetworkX’s Graph class stores undirected edges, making it suitable for UGs, but it carries no type-level semantic enforcement (no bidirected edges, no directed edges, etc.).↩︎

NetworkX has no mixed-edge graph class. Its four classes (Graph, DiGraph, MultiGraph, MultiDiGraph) are all homogeneous-edge. The pywhy-graphs extension provides mixed-graph types but is a separate package.↩︎

pgmpy has a single PDAG class (pgmpy.base.PDAG) documented as “also known as CPDAG”. It supports CPDAG construction from a DAG via DAG.to_pdag() (Chickering 2002) and to_cpdag() / apply_meeks_rules() methods. There is no separate dedicated CPDAG class, and PDAG does not enforce the invariant that the graph represents a complete Markov equivalence class.↩︎

No dedicated MPDAG class. PDAG.apply_meeks_rules(apply_r4 = True) applies all four Meek rules and could produce an MPDAG, but there is no type-specific MPDAG class or MPDAG-specific construction pathway.↩︎

AncestralBase stores mixed-edge graphs with arbitrary marks and is the parent of MAG. It is an internal base class, not a user-facing untyped mixed-edge graph class. Coverage of several typed mixed graphs (ADMG, MAG, PDAG) does not by itself satisfy this column.↩︎

Tetrad has no dedicated Cpdag class. CPDAGs are represented as EdgeListGraph objects containing directed and undirected edges. The library validates CPDAG status at runtime via Paths.isLegalCpdag() and produces CPDAGs through GraphTransforms.dagToCpdag() / GraphSearchUtils.basicCpdag(). The CPDAG concept is operationally first-class but not a distinct enforced type.↩︎

No dedicated Mpdag class. Tetrad validates MPAGs via Paths.isLegalMpag(), and MeekRules.orientImplied(graph) with a populated Knowledge object applies Meek rules under background constraints, but there is no single dagToMpdag factory.↩︎

No dedicated Mag class. MAGs are represented as generic EdgeListGraph objects and validated post-hoc via Paths.isLegalMag(). Constructed through GraphTransforms.dagToMag(dag) and converted to PAGs via GraphTransforms.magToPag().↩︎

No dedicated Pag class. PAGs are EdgeListGraph instances with circle, tail, and arrowhead endpoints. Validated via Paths.isLegalPag(). Produced by GraphTransforms.dagToPag() and GraphTransforms.magToPag().↩︎

SemGraph (in edu.cmu.tetrad.graph) enforces directed + bidirected edges only and automatically manages error nodes for each endogenous variable, implementing ADMG-like semantics. However, it is scoped to structural-equation-model graphs (with explicit error nodes) rather than a general-purpose ADMG class.↩︎

No dedicated undirected-graph class. Undirected edges can be added to EdgeListGraph via addUndirectedEdge(), and GraphUtils.undirectedGraph(g) converts all edges to undirected. There is no enforced UG type with invariant-checking.↩︎

EdgeListGraph is the universal mixed-edge container and accepts all five endpoint combinations (directed, undirected, bidirected, partially-oriented, nondirected). It does not commit to any single graph semantics, but it is not branded as a distinct “mixed/general” class—it is simply the default implementation of the Graph interface.↩︎

Parents and children are obtainable via neighbors(graph, v, mode = "in" / "out"), but there are no dedicated parents() / children() exported functions.↩︎

Ancestors and descendants are obtainable via subcomponent(graph, v, mode = "in" / "out"), but there are no dedicated ancestors() / descendants() exported functions.↩︎

No dedicated parents() or children() functions. inEdges() returns the set of in-neighbours (functional equivalent of parents on a directed graph); edges() returns out-neighbours (functional equivalent of children). Both are exported S4 generics.↩︎

acc() returns all nodes reachable from a given node on a directed graph, equivalent to the descendant set. There is no ancestors() equivalent exported by the package.↩︎

DFS() (exported) provides depth-first graph traversal; pathWeights() computes weights along a user-specified path. No all-paths enumeration, shortest-path, or directed-path-existence function is provided; those require the companion RBGL package.↩︎

parents() and children() are exported (R/graph-querygraph.R). There is no corresponding descendants() function in the exported API; ancestors() / ancestralSet() / ancestralGraph() cover the ancestor side but the symmetric descendent query is absent.↩︎

ancestors(), ancestralSet(), and ancestralGraph() are exported. A dedicated descendants() is absent from the exported API.↩︎

searchAM(amat, x, type = "pa" / "ch") returns parents and children for DAG, CPDAG, MAG, and PAG adjacency matrices. There is no standalone parents() or children() function; retrieval goes through the general searchAM() dispatcher.↩︎

searchAM(amat, x, type = "an" / "de") returns ancestors and descendants for DAG/CPDAG/MAG/PAG. possAn() and possDe() / possibleDe() return possible ancestors/descendants in PAGs. All operate on raw adjacency matrices via searchAM().↩︎

dsep(a, b, S, g) tests d-separation in a DAG (using a graphNEL object and moralization; Lauritzen 2004). dsepTest() wraps it for use as a CI oracle. There is no exported d-separation function for CPDAGs or PDAGs directly; dsep() is DAG-only.↩︎

dsepAM(X, Y, S, amat) tests m-separation on MAG adjacency matrices. dsepAMTest() wraps it for algorithmic use. Coverage is MAG-only; there is no exported m-separation function for PAGs in pcalg’s own API.↩︎

isValidGraph(amat, type = "dag") checks for directed cycles as part of DAG validation. There is no standalone is_acyclic() function; acyclicity is only surfaced through isValidGraph().↩︎

The skeleton() function estimates a skeleton from data using conditional independence tests (it is a causal discovery function). There is no separate exported function to extract or query the skeleton of an already-known graph.↩︎

udag2pdag() and dag2cpdag() identify v-structures as part of the CPDAG-completion process, but there is no standalone exported function to enumerate or check v-structures of a given graph.↩︎

No function enumerates v-structures as a set. isCollider(x, u, v, w) tests a single triple, but there is no exported vStructures() or equivalent. Skeleton (undirected version of the graph) is also absent as a dedicated function.↩︎

dsep() tests d-separation on bn objects. When the input contains undirected arcs (PDAG/CPDAG), the function calls cpdag.extension() to produce a single consistent DAG extension first; d-separation is then tested on that DAG rather than natively on the PDAG.↩︎

pa() returns parents and ch() returns children of a node set in a DAG. bd() returns the boundary (neighbours + parents). These operate on raw adjacency matrices and are undocumented for MAG/mixed-graph inputs; no descendants() or spouses() analogue is exported.↩︎

ancGraph() / anGraph() compute the transitive closure (ancestor relation) of a DAG as a Boolean adjacency matrix rather than returning a node list. No descendants() function is exported.↩︎

findPath() finds a single path between two nodes in an undirected graph (used internally by fundCycles()); it is not intended for direct user calls and does not enumerate all paths.↩︎

essentialGraph() returns the CPDAG (encoding the full MEC) but does not enumerate the Markov-equivalent DAGs. MarkEqRcg() / MarkEqMag() test pairwise Markov equivalence; RepMarDAG() / RepMarUG() / RepMarBG() find a single representative.↩︎

is_cyclic() is exported and checks directed cycles, and topologicalOrder() / isTopological() implicitly validate acyclicity, but is_cyclic() documentation notes it is “Not tested”—◐ reflects this caveat.↩︎

mb() finds the “Markov blanket for a vertex in an ancestral set” but requires the user to supply an explicit ancestral set A and the vertex must be childless in that set—a constrained helper rather than a standard graph-wide Markov-blanket function.↩︎

DiGraph.predecessors() and DiGraph.successors() return parent/child iterators. There is no dedicated parents() / children() API for a named causal DAG.↩︎

networkx.algorithms.d_separation provides is_d_separator, is_minimal_d_separator, and find_minimal_d_separator. All three require a DAG and raise NetworkXError on cyclic or undirected graphs. They do not support PDAGs, MAGs, or any other mixed-edge graph type.↩︎

NetworkX has no dedicated skeleton() function. The moral graph (nx.moral_graph) is available, but skeleton extraction (dropping directionality without moralization) is not an exported function.↩︎

DAG.get_ancestors() is implemented. There is no DAG.get_descendants() method; nx.descendants() from NetworkX can be called on the underlying nx.DiGraph directly, but it is not part of pgmpy’s exported API. ADMG does provide both get_ancestors() and get_descendants().↩︎

No skeleton() method exists in any pgmpy graph class. The undirected skeleton is accessible via the inherited nx.DiGraph.to_undirected(), but this is not a pgmpy-level exported function.↩︎

Paths.isMSeparatedFrom(x, y, z, isPag) implements d-/m-separation. The simpler overload’s Javadoc says “DAG only”, yet the isPag parameter explicitly switches PAG semantics. Per-graph-type documentation is inconsistent; users access d-separation as isMSeparatedFrom(..., false). The back-door helper Paths.isSatisfyBackDoorCriterion() constructs a Dag internally and throws if input is not a DAG.↩︎

There is no isAcyclic() method. Callers use !graph.paths().existsDirectedCycle() or Paths.isLegalDag() / GraphUtils.isDag(). Fully supported but requires negation of a cycle-detection predicate.↩︎

No dedicated skeleton() method. GraphUtils.undirectedGraph(g) converts all edges to undirected (producing the skeleton graph object), but it is a workaround.↩︎

GraphUtils.listColliderTriples(graph) returns all definite collider triples. However, EdgeListGraph.isDefCollider() checks arrowhead endpoints without testing non-adjacency of the outer nodes, so the method returns both shielded and unshielded colliders rather than v-structures (unshielded colliders) exclusively.↩︎

adjustment_set(cg, X, Y, type = "optimal") returns the O-set of Henckel/Perković/Maathuis (2019), but only for DAGs—the underlying Rust implementation errors on non-DAG graph classes. There is no O-set support for CPDAGs, MPDAGs, MAGs, or PAGs.↩︎

backdoor(amat, x, y, type) implements the Generalized Backdoor Criterion (GBC) of Maathuis & Colombo (2015), which subsumes Pearl’s classical back-door criterion. For DAGs the GBC reduces to Pearl’s back-door, so back-door is available as a special case, but the function does not specifically enforce or name the classical back-door separately. Generalized adjustment is adjustment(amat, amat.type, x, y, set.type) (full GAC, Perkovic et al. 2015/2018), and optimal is optAdjSet(graphEst, x.pos, y.pos) (Henckel/Perkovic/Maathuis 2019, DAG/CPDAG/PDAG only).↩︎

rmvnorm.ivent(n, object, target, target.value) simulates from an interventional distribution of a GaussParDAG by fixing target nodes — implements the do-operator semantics within the simulation framework but does not mutilate the graph object itself. There is no exported mutilate() or do() helper that returns a modified graph with incoming edges removed from intervention nodes.↩︎

adjustmentSets(effect = "total") and isAdjustmentSet() both invoke GraphTransformer.backDoorGraph() in the JS engine—Pearl’s back-door graph construction, not the generalized Perkovic criterion. The R documentation cites Perkovic et al. (2015), but the JS implementation (listMsasTotalEffect in jslib/graph/GraphAnalyzer.js) applies the back-door graph uniformly across dag/pdag/mag/pag. Generalized adjustment scores ○ because no separate generalized-adjustment code path exists. type = "canonical" returns the ancestor-based canonical set but is not an optimality criterion in the Rotnitzky/Henckel sense.↩︎

backDoorGraph(x) removes the first directed edge on every proper causal path (with PAG handling via pagToPdag conversion)—the back-door-graph helper, not a general do()-style mutilation that removes all incoming edges to an arbitrary intervention set. No mutilate() or do() function is exported.↩︎

mutilate() removes edges adjacent to a specified vertex set and supports a dir = -1 option that removes only incoming edges. It can mechanically produce the mutilated graph required by do(), but the function is documented as a general edge-deletion utility with no causal semantics, no treatment of bidirected edges in the do-calculus sense, and no convenience wrapper that enforces “remove all incoming edges to the intervention set.”↩︎

Paths.isSatisfyBackDoorCriterion(graph, x, y, z) implements Pearl’s back-door criterion check, but the implementation constructs new Dag(graph) internally and throws if the input is not a DAG. Back-door adjustment sets are not separately enumerated; the generalized RecursiveAdjustment.adjustmentSets() subsumes the back-door case for DAGs.↩︎

OSet.oSetDag(graph, X, Y) and OSet.oSetCpdag(graph, X, Y, maxPathLength) implement the Henckel–Perković–Maathuis O-set for DAGs and (amenable) CPDAGs respectively. The Ida and PdagPagIda classes wrap this as IDA_TYPE.OPTIMAL. Coverage is restricted to DAG/CPDAG; O-sets for MAGs/PAGs are not implemented.↩︎

to_dot() / write_dot() write Graphviz DOT format, but caugi has no DOT reader—round-trip via DOT is not supported.↩︎

caugi_serialize() / caugi_deserialize() (and read_caugi() / write_caugi()) read and write caugi’s own JSON-based serialization format (Rust-side: src/rust/src/lib.rs). It is not a standardized cross-package JSON schema.↩︎

DOT (Graphviz) format is supported for writing only (write_graph(format = "dot")). Reading DOT format is not supported. read_graph() accepts: edgelist, pajek, ncol, lgl, graphml, dimacs, graphdb, gml, dl.↩︎

igraph provides explicit coercion to/from graphNEL objects from the Bioconductor graph package via as_graphnel() and graph_from_graphnel(). No coercion helpers exist for dagitty, bnlearn, pcalg, or other causal R package object classes.↩︎

Write only. toDotWithRI() and toDotR() serialise graphNEL objects to DOT (Graphviz) format. No DOT reader is provided in the graph package.↩︎

Internal coercions are provided: as(graphNEL, "graphAM"), as(graphAM, "graphNEL"), as(graphBAM, "graphNEL"), as(matrix, "graphNEL"), as(matrix, "graphAM"), graph2SparseM() / sparseM2Graph() (requiring optional SparseM). No coercions to or from igraph, dagitty, or bnlearn objects are exported by graph itself.↩︎

gRbase provides as() S4 coercion methods that convert among igraph, graphNEL (Bioconductor graph), dense adjacency matrix, and sparse dgCMatrix representations. There are no explicit coercion helpers for other causal-graph packages (dagitty, bnlearn, pcalg).↩︎

pcAlgo and fciAlgo objects can be coerced to graphAM / graphNEL (from graph) via S4 coercion methods. The package wraps graphNEL as its primary DAG container, and iplotPC() converts to igraph for plotting. Incoming coercion (from igraph, bnlearn, dagitty, etc. into pcalg objects) is not provided.↩︎

convert(x, to = "igraph") and convert(x, to = "causaleffect") produce igraph objects (one-way: dagitty → igraph only). as.dagitty.bn() converts bnlearn bn objects to dagitty (one-way the other direction). lavaanToGraph() converts lavaan models to dagitty. No coercion to/from graphNEL, pcalg, or other classes.↩︎

write.dot() exports the network structure to Graphviz DOT format, but there is no corresponding read.dot()—DOT support is write-only.↩︎

grMAT() converts graphNEL (from graph) and igraph objects into ggm adjacency matrices. plotGraph() also accepts graphNEL/igraph as input. There is no function to export a ggm adjacency matrix back to a graphNEL or igraph object; coercion is one-directional (into ggm’s format).↩︎

DOT format read (read_dot) and write (write_dot) are available via networkx.drawing.nx_pydot, but pydot is an optional external dependency that must be installed separately. A pending deprecation warning has also been filed for nx_pydot (issue #5723).↩︎

NetworkX provides explicit conversion helpers to/from NumPy arrays, SciPy sparse arrays, and Pandas DataFrames. Conversion to/from other graph objects (igraph, graph-tool) is handled by those packages’ own from_networkx() / to_networkx() methods, not by NetworkX itself.↩︎

DAG.to_graphviz() returns a pygraphviz.AGraph from which DOT text can be obtained via .to_string(). There is no direct to_dot() / from_dot() API; reading DOT is not supported through pgmpy’s own API.↩︎

pgmpy graph classes (DAG, PDAG, ADMG) inherit from networkx.DiGraph/networkx.MultiDiGraph, so they are directly usable as NetworkX objects. There are no explicit coercion helpers to or from other graph-package objects (igraph, R dagitty objects, bnlearn objects).↩︎

GraphSaveLoadUtils.graphToDot(graph) writes Graphviz DOT format (string and file variants). There is no DOT reader, so round-trip is not supported. Write-only.↩︎

The C igraph library uses adjacency-list storage internally. On the R side, as_adjacency_matrix(sparse = TRUE) and graph_from_adjacency_matrix() accept/return sparse Matrix objects (dgCMatrix), but the igraph object itself is an opaque external pointer backed by the C library’s adjacency-list representation, not an R-level sparse matrix.↩︎

graphNEL uses an adjacency-list representation (sparse by nature). graphBAM (experimental) stores adjacency as a compact bit-array. graph2SparseM() / sparseM2Graph() convert to/from SparseM sparse matrices when that package is installed.↩︎

No built-in plotting. DOT serialisation helpers (toDotWithRI, toDotR) are provided, but actual rendering requires the separate Rgraphviz package. No layout algorithms are implemented in graph itself.↩︎

gRbase exports iplot() / graph-iplot, a thin wrapper around igraph’s plot.igraph(). There are no built-in layout algorithms: layout selection is delegated entirely to igraph.↩︎

pcAlgo and fciAlgo objects have plot() S4 methods that delegate to Rgraphviz. iplotPC() delegates to igraph’s plot. plotSG() also requires Rgraphviz. There are no native layout algorithms in pcalg itself.↩︎

graphLayout(x, method = "spring") implements a single spring/force-directed layout algorithm. The documentation explicitly states "currently, only 'spring' is supported".↩︎

The bn class stores arcs as a 2-column character matrix ($arcs) plus per-node children/parents lists in $nodes. The exported alst() function exposes an adjacency-list view. This is effectively sparse but is not a formal CSR/CSC or Matrix-class sparse type.↩︎

A rudimentary plot.bn() S3 method draws nodes in a circle with no external dependencies, but the package documentation describes it as “a last resort for when Rgraphviz is not available.” The full-featured graphviz.plot() requires the Bioconductor Rgraphviz package.↩︎

All graph layout computation is delegated to Graphviz via the Rgraphviz package (algorithms: dot, neato, twopi, circo, fdp); bnlearn implements no native layout algorithms.↩︎

plotGraph() and drawGraph() delegate layout to igraph (default layout.auto), so any igraph layout function can be passed via the layout argument. The layouts are not implemented natively in ggm.↩︎

Edge data can be stored in adjacency-list format (adjList class), which is a sparse representation. However, the package also supports dense adjacency matrices (adjMatrix class) and does not commit to a single internal format. No CSR/CSC or other documented sparse-matrix format is enforced.↩︎

plot.mixedgraph() is an S3 method but it delegates entirely to the pcalg / Rgraphviz stack by converting the graph to a fciAlgo object. Layout is determined by Rgraphviz; no layout algorithms are implemented within MixedGraphs. The function is listed under Suggests (pcalg, Rgraphviz), so plotting is unavailable without those packages.↩︎

The package exports a %G% infix operator that automatically converts a mixedgraph into the format expected by the right-hand-side function (from another package), enabling cross-package chaining. This is a narrow inter-package bridge operator rather than a full tidyverse-style verb chain that takes and returns mixedgraph objects.↩︎

NetworkX stores graphs internally as a dict-of-dicts-of-dicts (adjacency dictionary), which is effectively sparse for large, sparse graphs (only present edges are stored), but this is not a CSR/CSC/COO sparse matrix—it is a plain Python dictionary structure. Conversion to SciPy sparse arrays is available via to_scipy_sparse_array.↩︎

NetworkX includes nx.draw(), nx.draw_networkx(), and related functions that render graphs via Matplotlib (networkx.drawing.nx_pylab). Matplotlib is an optional dependency — NetworkX raises ImportError when it is absent. The drawing module is described in the documentation as basic/non-primary functionality.↩︎

DAG.to_graphviz() returns a pygraphviz AGraph (requires the optional pygraphviz dependency); DAG.to_daft() returns a daft PGM object. Neither is a plot() method that renders to screen directly. No built-in matplotlib-based plot() is present.↩︎

Layout computation is delegated to pygraphviz (via to_graphviz()) or to NetworkX spring-layout variants (inside to_daft()). pgmpy does not implement its own layout algorithms.↩︎

Tetrad’s GUI module (tetrad-gui) provides rich interactive graph visualization, but this is a Swing-based desktop application component, not a programmatic plotting API. From the library (tetrad-lib), LayoutUtil computes node coordinates (Kamada-Kawai, Fruchterman-Reingold, circle, square, causal-order layouts) and GraphSaveLoadUtils.graphToDot() exports to Graphviz for external rendering. There is no plot(graph) method in the library that renders directly to a file or screen without the GUI.↩︎