Skip to contents

Changelog

Source: NEWS.md

dgraphs (development version)

  • fermat.distances(return.graph = TRUE) now returns the union of selected shortest-path edges with powered base lengths and source/target coverage metadata, while preserving the default distance-matrix return.

  • Added fermat.knn() and create.fermat.sknn.graph() for exact complete-graph Fermat nearest neighbors using Euclidean pruning and truncated Dijkstra. Union graph lengths are path distances; global distance preservation and connectivity are not assumed.

  • Added an exact implicit complete-graph backend to fermat.distances(), with source-to-all and rectangular queries, a workspace estimate guard, and a teaching example supplying exact Fermat constraints to sparse metric MDS.

  • Extend the Fermat vignette with noisy arms and nested diffuse background, geometric path diagnostics and complete-versus-kNN distance checks.

  • Add fermat.distances() with a complete Euclidean graph reference, explicit symmetric-kNN or supplied-graph restrictions, and optional rooted output. Add a noisy-circle vignette comparing restricted distances to the reference.

  • Export create.ian.graph() with connectivity preservation and the audited retry-power numerical policy as defaults. Retain explicit strict/reference options, initial/final graphs, profile mapping and diagnostics; no fixed row cap.

  • Add build.ian.backend() for explicit optional-backend setup on macOS arm64. Backend compilation remains separate from ordinary package installation.

  • Geodesic nearest-neighbor queries now count other vertices, exclude self by identity, preserve coincident observations and break ties by vertex index. Insufficient reachable neighbors error unless small.component = "truncate"; truncation pads missing indices/distances with NA/Inf and reports effective.k.

  • Remove no-op intersection parallel controls, path.dist edge.col, geodesic.knnx method, plot.IkNNgraphs k, and the explorer run_dir alias and old cache option. graph.embedding verbose now reports elapsed layout time.

  • Add concise print/summary methods for geometry samples and path collections.

  • Include a deterministic saved circle benchmark in the explorer. Preserve coordinate proportions by default, label axes and optional normalization, and disclose edge thinning without changing saved metrics.

  • Make the package overview visible in installed help. Link every catalog and method entry to detailed help generated from the same installed Rd database, including offline copies. Add perspective static galleries and figure alt text.

  • Accept the additional binomial-generator field in R-devel RNG state headers when supported by the runtime, preserving replay tokens and existing word validation instead of rejecting freshly generated states.

  • Separate release/development installation examples, add a reproducible README figure and a pkgdown site, and validate static and interactive documentation separately in CI with exact source/archive fingerprints and Windows coverage.

  • Separate path measurement, geodesic neighbors, intersection stability and display modules; rename their regression suites by enduring workflow names.

  • Unify graph spectra across the native and R backends: nev counts the smallest positive eigenpairs in increasing order, and defaults to all positive modes. Results have class graph_spectrum, with exact component nullity, vertex component IDs, and a resolved zero-mode tolerance. All component zero modes are excluded. Counts are validated without truncation; empty/edgeless graphs return zero-column spectra. Both backends honor dense versus sparse Laplacian output; native full spectra use a dense solver.

  • Breaking spectral embedding interface: use graph.spectral.embedding(spectrum, dim, scale) with the complete spectrum object, replacing the separate eigenvector/eigenvalue arguments. Choose scale = "none" or "inverse.sqrt". Disconnected graphs must be embedded component by component. No compatibility wrappers are retained.

  • Correct subdivide.path() to include both endpoints and interpolate at equal arc-length intervals along the polyline, skipping repeated points instead of returning missing rows. Singleton and zero-length paths now have length zero, normalized cumulative distances of zero, and repeated-location subdivisions. Empty/malformed paths and invalid subdivision counts or vertex indices are rejected explicitly; subdivision requires at least two output points.

  • Add explore.graphs(), the Graph Reconstruction Explorer migrated from dggraphui, with optional Shiny dependencies. read.graph.benchmark() reads existing benchmark manifests; register.graph.project() remembers external projects without moving or changing their data.

  • Retain saved-layout viewing and add current weighted-GRIP support for explicit missing-layout generation. Cache keys distinguish full project paths.

dgraphs 0.3.0

  • Consolidate distance matrices under graph.geodesic.distances(), accepting either a graph or points with explicit symmetric-kNN/MST construction. MST construction preserves zero-length edges; k = 1 no longer implies an MST.

  • create.path.graph(graph, h.values) always returns a named collection. Correct hop-constrained route reconstruction and support reverse and self queries in get.shortest.path(). Remove the raw public shortest.path(), estimate.geodesic.distances() and create.path.graph.series() interfaces.

  • Replace plot2D.colored.graph() with plot() for dgraph. Layouts can be reused, numeric values and explicit colors are distinct, constant values plot correctly, and undirected edges are drawn once. Three-dimensional coordinates require explicit projection. Layout and plotting share stage selection. No compatibility wrappers are retained.

  • Replace ten basic graph constructors with create.graph(type, ...), without compatibility wrappers. Eight types share vertex counts, optional labels and strict argument validation. Chains use per-side span; circles return coordinates and explicit arc/chord lengths; random graphs enforce feasible edge budgets and support isolated seeds.

  • Breaking graph API: every constructor returns dgraph; use graph.adjacency, graph.lengths, graph.edges, graph.order, graph.stages and graph.edge.attribute. Raw metric inputs use adj.list and length.list. Old graph fields, conversion overloads, get.edge.weights and extract.edge.lengths are removed without compatibility wrappers.

  • Graph sequences use increasing k.values. Coordinate, geodesic and iterated intersection covers contain self plus k other vertices, with distance ties resolved by vertex index. Undersized components require explicit truncation. Neighbor caches use format 3 and validate working coordinates and row order.

  • Mutual-neighbor sequences now use the documented path/edge ratio directly; zero disables pruning. The former unintended addition of one is removed.

  • Graph layouts select edge attributes and transformations explicitly. Nerve overlaps are attributes rather than lengths. Path results contain $graph; path series compute each requested hop limit independently.

  • Expand quadratic-geometry examples with interactive ivue galleries, labeled x/y/z axes, all specialized sampling modes, frames, dimensions and metric diagnostics. Optional visualization dependencies have static fallbacks.

  • Add installed function-guide and synthetic-geometry vignettes, with a task-oriented catalog, method workflows, reproducible examples and explicit geometry/sampling limits. Link all guides from the package overview and README; add automated export and method-help coverage checks.

  • Fix one-point box sampling and multiple-height embeddings to retain matrix dimensions. These inputs previously failed in inherited geometry kernels.

  • Add explicit current-stream sampling for Geometry Lab compatibility with nondefault RNG kinds; the default isolated seed/state modes are unchanged.

  • Add reusable synthetic geometry constructors, embeddings, geometric edge lengths, and sampling specifications previously provided by geosmooth.

  • Add sample.synthetic.geometry() with explicit random-state continuation and restoration of the caller’s random-number state. Statistical responses, registry identities and the legacy G4 recipe remain in geosmooth.

  • Move the development Geometry Lab app and maintained quadratic-surface fixture/reference tools into this repository. Geometry Lab uses shared versioned sampling while retaining its existing point identities.

Graph construction and maintenance changes carried forward

  • Adds create.sknn.graphs() for constructing a sequence of symmetric-kNN graphs from one cached ANN search. create.sknn.graph() now also accepts a validated precomputed neighbor matrix and supports graph.detail = "minimal" for scalable fitting workflows that do not need repaired lifecycle branches.

  • Removed global compiler-warning suppression and unsupported native OpenMP branches. Native graph construction remains serial, including in custom builds with OpenMP enabled; compatibility arguments are retained.

  • ANN errors now propagate as R errors instead of silently continuing. Native entry points catch C++ exceptions, ANN point and tree ownership is exception-safe in nearest-neighbor and MST helpers, and the native kNN entry validates dimensions, finite coordinates and neighborhood size.

  • Documented all registered S3 methods and added runnable method examples. The documentation-coverage test now includes registered methods.

  • compare.adj.lists() is silent by default; use verbose = TRUE to print vertices whose neighbor sets differ.

  • Fixes summaries of empty detect.local.extrema() results by retaining the requested maxima/minima setting in a scalar detect.maxima component. Legacy empty objects without this metadata report an unknown (NA) type instead of being incorrectly labeled as minima.

dgraphs 0.2.0

CRAN release: 2026-08-31

Breaking API changes

  • Removes the deprecated create.radius.graph() and create.adaptive.radius.graph() compatibility wrappers. Use create.rknn.graph() with type = "fixed" or type = "adaptive.radius", respectively.
  • Removes the temporary public backend entry point cpp.create.rknn.graphs(). Use create.rknn.graphs(backend = "cpp").
  • Internalizes the low-level dist.to.knn(), euclidean.distance(), and graph.adj.mat() helpers.
  • Removes the superseded adjlist.to.igraph() conversion function. as_igraph() now accepts bare adjacency lists and an optional aligned weight.list argument.
  • Removes the unused graph-edit workflow comprising graph.edit.distance(), load.graph.data(), calculate.edit.distances(), and create.distance.plot().

Other changes

  • Fixes a C++ One Definition Rule violation reported by CRAN’s special link-time-optimization check. The minimum-spanning-tree implementation now uses a privately scoped edge type, and an unused conflicting intersection-kNN helper type has been removed. This does not change the R API or graph semantics.
  • Fixes an ANN fixed-radius boundary issue that could omit adaptive-radius edges lying exactly at a local-scale threshold. In particular, the adaptive maximum-radius rule with factor one now reproduces symmetric-kNN edges for exact, tie-free searches.
  • Clarifies that the historical rel_geodesic_stress diagnostic is a target-normalized graph-geodesic relative RMSE, not Kruskal’s Stress-1.
  • Adds a full workflow vignette covering graph construction, lifecycle diagnostics, connectivity repair, parameter sequences, conversion to igraph, and geodesic-isometry diagnostics.
  • Expands examples across the exported API in preparation for an R Journal package paper.
  • Extends as_igraph() to current dgraphs graph objects while preserving support for legacy basin graph objects.

dgraphs 0.1.0

CRAN release: 2026-08-20

  • First public release.
  • Provides mutual, shared-neighbor, intersection, geodesic nearest-neighbor, radius, adaptive-radius, and minimum-spanning-tree-completed graph constructors.
  • Provides graph conversion, weighting, pruning, diagnostics, spectral embedding, endpoint detection, and path utilities.
  • Includes native implementations for performance-sensitive graph construction and analysis.