Class SeedMarkovCheck

java.lang.Object
edu.cmu.tetrad.search.utils.SeedMarkovCheck

public final class SeedMarkovCheck extends Object
Harness drop-in: checks whether the graph FCIT starts from (the PAG of the BOSS/GRaSP DAG, i.e. G_0, BEFORE any edge removals) is a Markov PAG for the simulated true model.

Why this matters. FCIT's legality-and-revert guard and its inducing-path detector are both sound, yet a maximality violation can still reach the output if it was INHERITED from the seed rather than created by a deletion: the seed is never legality-checked (the check runs only after a removal), and FCIT only deletes edges, so it can never add back a true edge the seed is missing. A non-Markov seed therefore breaks the standing hypothesis of every theorem in the reachability section (prop:maximality Case 1 depends entirely on adj(G_0) >= adj(G*)), and the violation is outside what FCIT can fix.

Two checks, both against the known true DAG:

  1. Adjacency superset (decisive). Every adjacency of the true PAG G* = dagToPag(trueDag) must be present in the seed. A missing one is a true edge the seed dropped; since G* is maximal that pair is m-inseparable in the true model, so the seed claims a false independence (not an I-map) and bequeaths a permanent inducing path. Each missing edge is cross-confirmed by searching for a true separating set: finding none confirms a genuine defect; finding one would instead flag a dagToPag/G* artifact rather than a seed problem.
  2. Sound unshielded colliders (optional I-map half). Even with the right skeleton, a seed collider X*->Z<-*Y that is NOT a true collider lets the seed entail a false independence. For each unshielded collider in the seed we confirm the true model can separate X,Y by some observed set that EXCLUDES Z; if Z is forced into every separator, the collider is unsound.

Getting the seed: the cleanest hook is one line in Fcit.search() — right after this.pag = GraphTransforms.dagToPag(dag, ...) stash a copy (this.seedPag = new EdgeListGraph(pag);) and expose a getter. Failing that, reproduce it in the harness from the same search DAG you already build. Pass the true DAG with its latents typed NodeType.LATENT; this checker marginalizes them exactly as FCIT's pipeline does.

  • Method Details

    • check

      public static SeedMarkovCheck.Result check(Graph seedPag, Graph trueDag, Knowledge knowledge, int recursiveDepth, int sepsetMaxDepth, boolean checkColliders)
      Does the check.
      Parameters:
      seedPag - FCIT's starting PAG (G_0), before any removals.
      trueDag - the simulated true DAG, latents typed NodeType.LATENT.
      knowledge - knowledge used for dagToPag (pass FCIT's, or new Knowledge()).
      recursiveDepth - recursiveDepth used for dagToPag (FCIT passes its field; -1 is fine).
      sepsetMaxDepth - cap on conditioning-set size in the true-model sepset search (-1 = unbounded; keep modest, e.g. 6-8, for 20-node runs).
      checkColliders - run the optional sound-collider (I-map) half.
      Returns:
      The check result.