NEWS.md

In caugi: Causal Graph Interface

caugi 1.4.0

• Updates for compatibility with S7 > 0.2.2.

New Features

• Add components() to partition graphs into weakly connected components or strongly connected components of the directed part, including cyclic graphs of class "UNKNOWN".

Performance Improvements

• remove_edges() now removes edges directly in Rust, avoiding graph table reconstruction and redundant session cloning.
• Node and edge tables use faster constructors, reducing overhead when building graphs and retrieving nodes() and edges().

Bug Fixes

• mutate_caugi() no longer forces simple = TRUE on every target class. A non-simple graph converted to "UNKNOWN", "ADMG" or "AG" keeps simple = FALSE, so e.g. a bow (A %-->% B, A %<->% B) converts to a valid ADMG instead of a simple graph with parallel edges that failed on build() or any query. The converted graph is now built eagerly, so invalid conversions fail immediately.

• m_separated() now handles ADMG collider paths with multiple bidirected edges, such as a -> b <-> c <-> d given {b, c}. This also fixes is_valid_adjustment_admg() and all_adjustment_sets_admg(), which use the same separation routine in the proper backdoor graph.

• dag_from_pdag() no longer adds v-structures the input PDAG did not have. Dor and Tarsi's condition (b) requires every undirected neighbour of the candidate sink to be adjacent to all its other neighbours; the previous implementation only required the undirected neighbours to be pairwise adjacent, so on e.g. A --> B <-- C with B --- D it oriented D --> B and created two v-structures. The algorithm now lives in Rust, like every other graph algorithm in the package.

• == and != for caugi objects now compare a symmetric edge as an unordered pair, so caugi(A %---% B) == caugi(B %---% A) is TRUE as it should be. This affected ---, <-> and o-o, and any symmetric edge added with register_caugi_edge(); whether a glyph is symmetric is read from the graph's own registry rather than hard-coded. Asymmetric glyphs remain direction-sensitive. testthat's expect_equal() now agrees with ==.

• Graphs now keep the edge definitions they were built with after reset_caugi_registry() and re-registration. Previously ==, add_edges(), mutate_caugi() and other verbs, write_caugi(), caugi_serialize() and to_graphml() used the current global registry instead, so they could compare edges wrongly, fail with Invalid edge code, or write the wrong glyphs. Adding an edge type the graph was not built with is now an error.

caugi 1.3.0

New Features

• Add enumerate_dags() to enumerate every DAG in the Markov equivalence class of a PDAG, and count_dags() to return the MEC size without materializing every DAG (#297).
• plot() now automatically bends edges around non-incident nodes that they would otherwise pass straight through, so edges between collinear nodes (e.g. within a tier) and edges crossing unrelated nodes stay visible. This is controlled by edge_style$route (default TRUE); disable it with edge_style = list(route = FALSE), or per edge type/edge via the usual edge_style overrides.
• Add == and != methods for caugi objects so cg1 == cg2 returns a single logical comparing graph content (class, nodes, edges, simple) rather than session identity.
• Add first-class "CPDAG" graph class support across the constructor (caugi()), coercion (as_caugi()), and class mutation (mutate_caugi()). Construction validates the full CPDAG invariant (chordal chain components, an acyclic component DAG, Meek closure, and strong arrow protection). generate_graph(class = "CPDAG") now returns a graph with @graph_class = "CPDAG" instead of "MPDAG", the precise label for the essential graph of a Markov equivalence class. Predicates defined on PDAGs and MPDAGs (is_pdag(), is_mpdag(), etc.) continue to accept CPDAGs unchanged.
• Meek-closed PDAGs are now reported with @graph_class = "MPDAG" instead of "PDAG". This affects the result of meek_closure() and generate_graph(class = "CPDAG"). Predicates and verbs defined on PDAGs (is_pdag(), mutate_caugi(), etc.) continue to accept MPDAGs unchanged.

Improvements

• aid() is now implemented natively in caugi's Rust backend, removing the external gadjid dependency (and its pinned git revision and vendored sources). Results are unchanged. The AID implementation (src/rust/src/graph/aid.rs) is a derivative of gadjid and is licensed MPL-2.0; the rest of the crate remains MIT. aid() now takes inputs of class "DAG" or "CPDAG" (previously "DAG" or "PDAG").
• Licensing of bundled code is now declared for CRAN: a top-level LICENSE.note documents the MPL-2.0 component (the AID files derived from gadjid) and the licenses of the vendored Rust crates, and their copyright holders are recorded via cph/ctb roles in Authors@R.
• adjustment_set(type = "backdoor") now returns an inclusion-minimal backdoor adjustment set, computed in linear time as a minimal d-separator in the proper backdoor graph, rather than the full set of parents of the exposure.
• The performance vignette is now a true vignette (rather than a pkgdown-only article) and is rebuilt manually from a cross-language harness under tools/benchmark/. The harness compares caugi to igraph, bnlearn, dagitty, ggm, pcalg, pgmpy, and Tetrad on a n ∈ {100, 1000, 10000} grid. The d-separation benchmark now uses a minimal d-separator computed via minimal_separator() (previously used a backdoor adjustment set, which is not in general a d-separating set).

Bug Fixes

• Fix hd() returning results that depended on the order in which nodes were declared. The Hamming distance now aligns nodes by name before comparing, so logically identical graphs always give the same distance (#323).
• Fix dag_from_pdag() failing with `from`, `edge`, `to` must be equal length. when a sink had multiple undirected neighbors (#298).
• Fixed a bug causing a partially undirected (--o) edges to be plotted as undirected edges.
• Fix adjustment_set(type = "backdoor") returning an invalid (often empty) set when a parent of the exposure lies on a backdoor path but is not an ancestor of the outcome (#308). The result is now always a valid backdoor adjustment set.
• Fixed is_mag() returning incorrect results for some ancestral graphs (#309). Adjacency was tested by binary-searching the concatenation of separately sorted neighbor buckets, which is not globally sorted, so some adjacent pairs were missed.
• Stripped non-build files (tests/, examples/, trybuild fixtures) from the vendored Rust dependencies so no vendored path exceeds 100 characters. This silences pak's "very long paths" warning and avoids installation failures on Windows without long-path support (#319).
• Fixed to_dot() and to_mermaid() (and write_dot()/write_mermaid()) silently converting partial --o and o-o edges into plain directed edges, dropping the circle endpoints (#307).

Deprecations

• The first argument of dag_from_pdag() is now named cg, matching the convention used by the rest of the package. The previous name PDAG continues to work as an alias but emits a deprecation warning.

caugi 1.2.0

New Features

• Add caugi_layout_circle() and a "circle" method for caugi_layout() that places nodes evenly along the perimeter of a circle (#108).
• Add list_caugi_edges() function to list all available edge types.
• Add first-class "MPDAG" graph class support across constructor, class mutation, and class resolution. class = "AUTO" now resolves Meek-closed PDAGs to "MPDAG".

Improvements

• Improved performance of all queries. Speedups are more significant on larger graphs, but even on small graphs, queries are roughly 5x faster.
• exogenize() is now implemented in Rust for DAGs, which reduces overhead on larger graphs.
• normalize_latent_structure() is now implemented in Rust for DAGs for faster latent normalization workflows.
• minimal_d_separator() is renamed to minimal_separator() and now supports ADMG and AG inputs (previously DAG-only), returning a minimal m-separator. Implemented via the unified linear-time algorithm of van der Zander & Liśkiewicz (UAI 2020). The old name minimal_d_separator() remains as a deprecated alias.

Bug Fixes

• Fix m_separated() on ADMGs: moralization now marries every pair in pa(v) ∪ sp(v), not just pa(v). The old code missed moral edges from bidirected co-parents and gave false positives (e.g. claimed Z ⊥ Y | X for Z -> X -> Y, X <-> Y).
• Fix is_valid_adjustment_admg() and all_adjustment_sets_admg() to verify the GAC's m-separation condition in the proper backdoor graph rather than via a per-neighbour decomposition. The old check trivially accepted neighbours of X that were themselves in Z, so it falsely classified {C} as a valid adjustment set in the M-bias ADMG C -> X, C <-> X, C -> Y, C <-> Y, X -> Y (#277).

caugi 1.1.0

New Features

• Add normalize_latent_structure(), which normalizes the latent structure of a DAG while preserving the marginal model over observed variables.
• Add minimal_d_separator(), which computes a minimal d-separator between sets of nodes in a DAG, with support for mandatory inclusions and restrictions.
• Add posteriors() query function, which is the dual of anteriors(). It returns all nodes reachable by following paths where every edge is either undirected or directed away from the source node. For DAGs, posteriors() equals descendants(). For PDAGs and AGs, it includes both descendants and nodes reachable via undirected edges.
• You can now specify whether to use an open or closed graph definition for the queries ancestors(), anteriors(), descendants(), and posteriors(). This can be set globally with caugi_options() or locally with the open = TRUE/FALSE argument. The default remains open = TRUE.
• Add is_mpdag() query to check whether a PDAG is closed under Meek's orientation rules (R1-R4), and meek_closure() to orient all implied edges until Meek closure.

Improvements

• caugi_options() now supports nested key drilling: multiple unnamed arguments traverse nested options (e.g., caugi_options("plot", "tier_style", "fill")).
• Rust remains the single source of truth for graph state. Graph properties (simple, graph_class, nodes, edges) are sourced from the session.
• A session is always created, including empty graphs (n = 0), which simplifies property access.
• Deprecated compatibility properties @.state, @name_index_map, @built, and @ptr now warn on access and return NULL.
• Deprecated compatibility constructor arguments build and state in caugi() now warn and are ignored.
• The inplace parameter in verb functions (add_edges(), remove_edges(), set_edges(), add_nodes(), remove_nodes()). All graph modifications now use copy-on-write semantics for consistency with R conventions. The parameter is deprecated and ignored with a warning.
• Added all.equal and compare_proxy methods for caugi objects to support graph-content comparison in tests.
• Add asp parameter to plot() for controlling the aspect ratio. When asp = 1, the plot respects equal units on both axes, preserving the layout coordinates. Works like base R's asp parameter (y/x aspect ratio) (#195).
• Add pdag_to_dag() function that generates a random DAG consistent with a given CPDAG/PDAG structure if possible (#201).

Bug Fixes

• Fixed a bug causing plot() to use incorrect layout if node names were not in the same order as in the graph object (#198).
• Fixed set_edges() so that it correctly replaces symmetric edges in simple graphs.

Deprecations

• The parameter all in districts() has been deprecated. Use districts() without arguments to get all districts.

caugi 1.0.0

New Features

• Add support for Ancestral Graphs (AGs), which combine directed (-->), bidirected (<->), and undirected (---) edges while satisfying ancestral graph constraints. New functions: is_ag(), is_mag().
• Add mode argument to neighbors()/neighbours() to filter neighbors by edge direction or type ("all", "in", "out", "undirected", "bidirected", "partial"). This is a structural query, and not a semantic query!
• neighbors() now supports class = "UNKNOWN" graphs, including mode-based filtering.
• Add simulate_data() that enables simulation from DAGs using SEMs. Standard linear Gaussian SEMs are defaults, but more importantly custom SEMs are available.
• Add "AUTO" parameter for class in caugi objects. This automatically picks the graph class in order DAG, UG, PDAG, ADMG, AG.
• Add Ancestral Graphs (AG) with support for directed, bidirected, and undirected edges, plus new queries is_ag() and is_mag() and m-separation for AGs.
• Add exogenize() function that exogenizes variables for any graph type. Current implementation is written in R, but it is so simple that it might be preferable over a Rust implementation. This might be changed later.
• Add latent_project() function that does latent projection from DAGs to ADMGs.
• Add native caugi serialization format for saving and loading graphs. New functions: write_caugi(), read_caugi(), caugi_serialize(), and caugi_deserialize(). The format is a versioned JSON schema that captures graph structure, class, and optional metadata (comments and tags).
• Add plot() method for visualizing graphs using various layout algorithms. The plot is rendered using grid graphics and returns a caugi_plot object that can be customized with node_style, edge_style, and label_style arguments. The plot() method accepts layouts as strings, functions, or pre-computed data.frames.
• Add caugi_layout() function to compute node coordinates for graph visualization.
• Add dedicated layout functions: caugi_layout_sugiyama(), caugi_layout_fruchterman_reingold(), caugi_layout_kamada_kawai(), caugi_layout_bipartite(), and caugi_layout_tiered(). Each function provides an API for its specific algorithm.
• Add to_dot() and write_dot() functions for exporting caugi graphs to DOT (graphviz) format. The resulting object is a new S7 class, caugi_export, which has a knit_print() method for rendering DOT graphs in R Markdown and Quarto documents.
• Add GraphML and Mermaid import/export support: to_graphml, write_graphml, read_graphml, to_mermaid, write_mermaid, and read_mermaid
• Add plot composition operators for creating multi-plot layouts: + and | for horizontal arrangement, / for vertical stacking. Compositions can be nested arbitrarily (e.g., (p1 + p2) / p3).
• Add caugi_options() function for setting global defaults for plot appearance, including composition spacing and default styles for nodes, edges, labels, and titles.
• Add caugi_default_options() function to query or reset to package default options.
• Add a new vignette, "Graph Visualization with caugi", demonstrating the new plotting capabilities and customization options.

Improvements

• Add favicons for the package website.
• Standardize is_caugi() validation calls internally.
• Adopt air as the R code formatter for the package.
• caugi_layout_tiered() now returns a tier column and orientation attribute in the layout data.frame, allowing plot() to automatically use tier information without requiring the tiers argument to be passed again.

Bug Fixes

• Fix typo in error messages and documentation examples.
• Remove unused index_name_map parameter from internal .cg_state() function.
• Fix interpretation of directed edge endpoint positions in the Rust backend, improving correctness of direction-dependent algorithms/metrics (e.g. topological sorting and structural Hamming distance computations).
• Fix bug in is_cpdag function that returns TRUE on non-complete PDAGs.
• Fix bug in shd returning positive values for equivalent graphs given in shuffled order.

caugi 0.4.0

• Add support for Acyclic Directed Mixed Graphs (ADMGs), which combine directed edges representing causal relationships with bidirected edges representing latent confounding.
• Add new functions for querying ADMGs: is_admg(), spouses(), districts(), and m_separated() (generalization of d-separation for graphs with bidirected edges).
• Add functions for adjustment set validation in ADMGs: is_valid_adjustment_admg() and all_adjustment_sets_admg() implementing the Generalized Adjustment Criterion.
• Add mutate_caugi() function that allows conversion from one graph type to another.
• Add custom printing method for caugi objects.
• Add optional edges_df argument to caugi() for easier construction from existing data frames containing the columns from, edge, and to.
• Improve error handling across all graph types (DAG, PDAG, UG, ADMG) with more descriptive error messages.
• Update as_adjacency() and as_igraph() to support bidirected edges.
• Update as_caugi() documentation to include "ADMG" as a valid class type for conversion.

caugi 0.3.2

• Change website to caugi.org/.
• Minor modifications to CONTRIBUTING.md.
• Minor README rewrite.

caugi 0.3.1

• Remove the use of lockBinding and unlockBinding in the package to silence R CMD check notes.

caugi 0.3.0

• Add mutate_caugi function that allows conversion from one graph type to another.
• Add custom printing method.
• Add optional edges_df argument to caugi for easier construction from existing data frames containing the columns from, edge, and to.
• Update How to use caugi in a package vignette to use new conversion functionality.
• Add CONTRIBUTING.md to github.

caugi 0.2.1

• Update function documentation to make package CRAN ready.
• Update performance vignette and change it to article.
• Add Michael Sachs and Johan Larsson to Authors in DESCRIPTION.
• Patch S4 class reading for as_caugi.

caugi 0.2.0

• Drop dependencies on dplyr and tibble.
• Improve performance by letting all data wrangling be done by data.table.
• Edges and nodes are now data.tables.

caugi 0.1.0

• Add Undirected Graphs (UG) to caugi.
• Refactor Rust backend for DAGs.
• Add NEWS.md!

caugi documentation built on Oct. 28, 2026, 1:07 a.m.