Continuous-time Bayesian network
Model Families AIMech:ContinuousTimeBayesianNetwork DRAFT
A finite-state continuous-time model family combining a separate initial Bayesian-network distribution with a directed, possibly cyclic graph of current-parent-dependent conditional intensity matrices, whose single-variable jump rates amalgamate into a homogeneous joint Markov generator.
Identity Notes & Concept Scope
Scoped to the finite homogeneous CTBN model of Nodelman, Shelton and Koller: an arbitrary separate initial BN and a possibly cyclic current-parent CIM dynamics graph. This is distinct from hidden/observed emission-model variants, discrete memory mixtures, undirected static potentials, timestamp-density topic models and continuous-state linear layers. The pinned BayesianNetwork DAG does not identify the dynamics graph; MarkovChain supplies a qualified process association. The source dossier read 127 selected definitions, 11 exact AIO node/incident/generated contexts and 11 full native neighbors; root and supporting reading scopes are separately retained. Before this creation all 442 AIO and 713 preceding native IDs, names, aliases and all scoped synonyms were refreshed and four complete checkouts searched including ignored and hidden files; no bounded identity collision was identified. The existing CTBN frontier owner is updated in place, without declaring inference/learning extensions exhausted.
Claims & Grounded Evidence
6 Claims
Mechanism
#initial-law-and-local-intensities
Specify finitely many finite-state variables, an initial joint distribution represented by an ordinary Bayesian network, and a separate directed dynamics graph. The dynamics graph may have cycles. For every variable X and current parent assignment u, specify a conditional intensity matrix Q_(X|u): off-diagonal entries are finite nonnegative transition rates, and each diagonal is the negative sum of that row’s outgoing rates. A parent change can select a different local matrix. The joint model is homogeneous because these rates are fixed functions of the current complete state, rather than explicit functions of clock time. The initial-law graph and dynamics graph have different roles and need not be the same.
Mechanism
#single-coordinate-amalgamation
For distinct joint states x and y differing only in coordinate k, set Q(x,y) to the local rate for x_k to y_k under the parent values in x. Set Q(x,y)=0 if more than one coordinate differs, and set Q(x,x)=-sum_(y!=x) Q(x,y). This assembles the local generators without multiplying their entries. The resulting finite matrix has nonnegative off-diagonal entries and zero row sums. The absence of a single jump changing multiple variables is a defining restriction of this construction; continuous time alone does not impose it on every Markov process.
Mechanism
#event-rates-and-zero-exits
While the complete state x remains fixed, let lambda=-Q(x,x). If lambda>0, its next-event waiting time is exponential with rate lambda, and destination y has probability Q(x,y)/lambda. This can be factored into choosing a variable in proportion to its local exit rate and choosing its new value in proportion to that local row’s off-diagonal rates. A variable with current exit rate zero contributes no event and must not divide by that rate. It may become enabled after a parent changes. If the total joint exit rate is zero, the joint state is absorbing. These guards qualify the printed destination/event formulas; the paper does not supply a guarded implementation for every zero-rate case. A child’s unconditional holding time need not be exponential at its initial local rate when its parents can change first.
Mechanism
#finite-interval-transitions
For the joint homogeneous generator, the transition matrix over duration delta is exp(delta*Q), and a row-vector distribution propagates as p(t)=p(s)*exp((t-s)*Q). Two-time joint probabilities instead have cells p_i(s)*[exp((t-s)*Q)]_ij; the source’s row-vector product labeled a two-time joint law is only the later marginal. A zero multi-coordinate generator entry does not forbid sequential changes during a positive interval. For two independent irreversible binary switches with rates 1 and 2, Q(00,11)=0 but P(00 to 11 in time t)=(1-exp(-t))*(1-exp(-2*t))>0 for t>0. This is a curator mathematical example, not a reported empirical result.
Limitation
#initial-dependence-boundary
The dynamics graph alone does not establish independence of complete variable trajectories under the arbitrary initial BN allowed by the definition. A curator counterexample uses two binary variables with no dynamics edges, rate-one flips in both directions, and initial law (3/8,1/8,1/8,3/8) in state order (00,01,10,11). Already P(00)=3/8 differs from the marginal product 1/4. At time t the joint law is (1/4,1/4,1/4,1/4) plus exp(-4*t)*(1/8,-1/8,-1/8,1/8), so the zero-state indicators have covariance exp(-4*t)/8 at every finite t>=0. Thus the unqualified trajectory-independence statement in Theorem 3.9 is not adopted. No replacement general graphical-independence theorem is asserted.
Scope
#representation-and-inference-scope
The local rate factorization can represent the model compactly, but its joint state space has the product of the local state cardinalities. Marginalizing variables need not leave a homogeneous Markov process on the remaining variables; parent trajectory information can matter beyond a one-time parent marginal. Sparse instantaneous dependencies also need not remain sparse over a nonzero observation interval. The paper’s clique-tree approximation, linearized marginalization and subsystem approximation are inference choices, not defining parts of this model family. No claim that every equivalent finite-time DBN is fully connected, or that the model guarantees cheap exact inference, is made.
⚠️ Curated Limitations & Negative Results
- The maintained scope is finite variables and finite, time-independent conditional rates with one local transition owner per variable. Time-dependent rates, infinite state spaces and simultaneous multi-variable jump extensions require separate definitions. Zero rates and reducibility are permitted; irreducibility and convergence to a unique stationary law independently of the initial distribution are not guaranteed.
- This is a stochastic model representation, not a particular simulator, learning procedure or approximate-inference algorithm. Hidden variables, emission distributions, action/reward control and intervention semantics are not required by this definition.
- Ancillary displayed inconsistencies in the corrected source remain outside the core construction: Example 2.2 on PDF p. 2 reverses two probabilities in the third embedded-chain row; the first marginalization display on p. 8 mixes its initial/destination eliminated-state indices; and Section 5.2.2 on p. 9 labels an expected holding time as an intensity, although its example uses the reciprocal. No repaired approximate-inference implementation or author-issued erratum is claimed.
- The author-hosted PDF is labeled Corrected without an established correction date or comparison to the original proceedings bytes. Original finite calculations support bounded examples; nonrational exponential probabilities are enclosed by rational bounds, not represented as exact rational values. No author-code execution, experiment replication, clinical validation or universal performance claim is included.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:MarkovChain
Markov Chain
|
related_match | The complete finite-state model defines a continuous-time Markov process, and the pinned target includes Markov Process as an exact synonym. This related mapping does not equate the specific CIM factorization with the broad AIO node or assert inheritance of its Network classification or ProbabilisticHiddenLayer has-part relation. The separate initial BN does not make the possibly cyclic dynamics graph an AIO BayesianNetwork DAG. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
nodelman2002
|
Uri Nodelman, Christian R. Shelton, Daphne Koller
|
primary paper
full_text
|
sha256:27c0afd16350…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-15T00:11:24Z
Updated: 2026-10-15T00:11:24Z
Source File:
data/curated/ContinuousTimeBayesianNetwork.yaml