Filtration (Probability Theory)¶
An ordered family of nested event σ-algebras on one probability space that represents the information available at successive indices.
Core Idea¶
A probability filtration is a family \((\mathcal F_t)_{t\in T}\) of sub-σ-algebras on one probability space, ordered so that \(s\leq t\) implies \(\mathcal F_s\subseteq\mathcal F_t\). It formalizes which events are measurable by each index; the same information may persist unchanged between indices. A stochastic process can be adapted to this family, but no process, martingale, stopping rule, completion or right-continuity condition is needed for the bare filtration identity.[ref-a32313dcd1bc][ref-1cf3fc5837b7]
The natural filtration of a process is its smallest history-generated instance. A larger family can still be a filtration and can matter to the process's conditional or independence claims. Thus general Filtration is distinct from live Natural Filtration, and from the homonymous algebraic Filtration whose stages are algebraic subobjects rather than event σ-algebras.[ref-a32313dcd1bc][ref-80164130d4d7]
Scope of Application¶
In discrete time, MIT's \(\mathcal F_k=\sigma(X_0,\ldots,X_k)\) records a sequence of observations. Its derived process \(Y_k=\operatorname{sign}(X_0+\cdots+X_k)\) is adapted, though the full history \(\mathcal F_k\) contains more event information than \(Y\)'s signs alone. In continuous time, Berkeley describes a Poisson count process relative to \(\mathcal F_t\); that family is often generated by the count history but need not be. The two settings share common carrier, ordered stages and inclusion, not a fixed time grid or compulsory natural-history choice.[ref-80164130d4d7][ref-1cf3fc5837b7]
Clarity¶
\(A\in\mathcal F_t\) means event \(A\) is resolvable from the modeled information at \(t\), not that \(A\) actually occurred. An adapted process satisfies a measurability relation to a filtration; a martingale adds a conditional-expectation relation; a Poisson-process definition adds distribution and increment-independence conditions. None should be silently substituted for the underlying nested-family definition. Nor does inclusion require strict information growth at every step.[ref-a32313dcd1bc][ref-80164130d4d7][^ref-1cf3fc5837b7]
Manages Complexity¶
The family condenses observation histories into a mathematically testable information state at each index. It lets one ask whether a value is measurable now and what a conditional expectation is given now, without enumerating every observation. But choosing a richer or poorer filtration may alter the answers: calling the sign process's ambient history its own natural history loses the distinction MIT explicitly draws. Extra regularity may be necessary for a particular theorem even though it is not constitutive of the bare object.[ref-80164130d4d7][ref-a32313dcd1bc]
Abstract Reasoning¶
Fix \((\Omega,\mathcal F,P)\) and an ordered \(T\). A filtration is an order-preserving assignment \(t\mapsto\mathcal F_t\) into the sub-σ-algebras of \(\mathcal F\) under inclusion. A constant family is valid. If \(X\) is adapted, its own natural family \(\mathcal N_t^X=\sigma(X_s:s\leq t)\) lies within the chosen \(\mathcal F_t\), but equality is not assured. For a Poisson count \(N_t\), the count-generated choice \(\sigma(N_s:s\leq t)\) is one such family, not the only possible one.[ref-a32313dcd1bc][ref-80164130d4d7][^ref-1cf3fc5837b7]
Knowledge Transfer¶
For another stochastic model, specify its common probability space, time index, event σ-algebras and inclusion law before attaching a process or theorem. Then separately ask whether a process is adapted, whether the family is natural for it, and which regularity or conditional assumptions the intended result requires. Transfer those formal roles from discrete increments to continuous arrival counts, not the same observation source or application-specific interpretation. The proposal is staged unparented: live Natural Filtration is narrower, Algebraic Filtration a distinct homonym, and portable Order/Information primes do not yet establish a necessary typed DAG edge.[ref-a32313dcd1bc][ref-80164130d4d7][^ref-1cf3fc5837b7]
[^ref-a32313dcd1bc]: Hao Wu, “Lecture 15: Introduction to martingales,” MIT OpenCourseWare 18.445, 8 April 2015, slide 10/PDF p. 9. https://ocw.mit.edu/courses/18-445-introduction-to-stochastic-processes-spring-2015/fb97374ded39d0e4a0047564af28dc93_MIT18_445S15_lecture15.pdf [^ref-80164130d4d7]: Yury Polyanskiy, “Lecture 25: Martingales I,” MIT OpenCourseWare 6.436J/15.085J, 2018, PDF p. 2 for generated history and sign-process example; classroom-use notes. https://ocw.mit.edu/courses/6-436j-fundamentals-of-probability-fall-2018/d734a5f3c245eef581b2937b53d8e6dd_MIT6_436JF18_lec25.pdf [^ref-1cf3fc5837b7]: Jim Pitman, scribe Jonathan Weare, “Lecture 26: Processes with Independent Increments,” Berkeley Stat205A, Fall 2002, PDF pp. 1–2 for filtration-relative Poisson process and count construction. https://www.stat.berkeley.edu/~pitman/s205f02/lecture26.pdf
Neighborhood in Abstraction Space¶
Filtration (Probability Theory) sits in a sparse region of the domain-specific corpus (61st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Codes, Matrices & Combinatorial Problems (30 abstractions)
Nearest neighbors
- Arithmetic Progression — 0.85
- Filtration (algebra) — 0.85
- Self-Organizing List — 0.85
- List (computing) — 0.85
- Big O in probability notation — 0.84
Computed from structural-signature embeddings · 2026-10-08