Affine Process¶
A Markov process whose conditional transition characteristic function is exponential-affine in its initial state.
Core Idea¶
An affine process is a Markov process whose transition distribution has an exponential-affine characteristic transform in the initial state. On a stated admissible state space \(D\), the foundational definition writes, for a horizon \(t\) and Fourier argument \(u\),
Here \(\phi\) and \(\psi\) may depend on \(t\) and \(u\), but not on \(x\); the condition holds for the stated family of starting states and arguments, not just for one fitted trajectory. The displayed expectation is the conservative-process case; when killing is allowed, use the characteristic transform of the possibly subprobability transition kernel, equivalently inserting survival before \(t\) in the expectation. Duffie, Filipović and Schachermayer formulate this on \(D=\mathbb R_+^m\times\mathbb R^n\) and also discuss more general admissible state spaces. It is the conditional transition transform, not merely a linear-looking stochastic differential equation or a financial use, that identifies the class.[1]
The frozen Wikipedia provenance, Basic affine jump diffusion, is narrower. An affine jump diffusion is one important regular realization; jump paths, particular jump-size laws, short-rate interpretations and pricing formulas are not necessary for the broader identity.[2][1]
Structural Signature¶
Sig role-phrases:
- Markov transition family. A present state \(x\) in a declared state space determines a transition law at horizon \(t\). Without this current-state conditional law, the transform to be classified is absent.[1]
- State-wide comparison frame. The claim ranges over initial states \(x\), eligible horizons \(t\) and Fourier arguments \(u\). One observed path, state or numerical transform point cannot establish the identity.[1]
- Exponential-affine dependence. The characteristic transform is \(\exp(\phi+\langle\psi,x\rangle)\), with the logarithmic exponent affine in \(x\). Without this condition the process can still be Markov, but not affine in this technical sense.[1]
These three roles are constitutive. Additional regularity can yield affine generator coefficients and generalized Riccati equations; those are qualified characterizations, not an extra role forced into the bare definition.[1]
What It Is Not¶
“Affine” does not mean that every sample path is a straight line, that the marginal density is affine, or that the SDE drift alone is affine. It is a condition on an entire family of conditional characteristic functions across starting states. The state space and admissible parameters matter: an algebraic formula that does not define a valid transition law on \(D\) is not enough.[1]
It is not synonymous with basic affine jump diffusion. A continuous Gaussian Ornstein–Uhlenbeck (OU) model is an affine-process case with no jumps, while the narrow jump-diffusion literature adds particular drift, covariance and jump-intensity forms. Nor is every affine process a bond-pricing or credit model.[1][2]
Scope of Application¶
The foundational regular theory connects OU-type processes on real-valued coordinates and continuous-state branching-with-immigration processes on nonnegative coordinates. The commonality is the transform's dependence on the initial state, despite different path behavior and state-space geometry. A square-root Cox–Ingersoll–Ross (CIR) diffusion is a familiar nonnegative branching-type example, while Gaussian OU mean reversion provides a real-valued example.[1][2]
The class has major financial uses because a tractable state transform can feed a model of rates, default intensity or stochastic volatility. But discount factors, time-integral transforms, survival probabilities and option prices require added rate/intensity/payoff definitions and suitable existence conditions. The original source treats such applications in separate sections; they should not be read backward into the definition.[1][2]
Clarity¶
There are two distinct senses of “affine” in play. The defining sense is an exponential-affine transition transform. For regular affine processes, the foundational characterization describes state-affine generator coefficients and generalized Riccati evolution of transform coefficients. Regularity includes stochastic continuity and a derivative condition; the source explicitly gives an affine Markov example outside that regular characterization. Therefore an entry test based solely on a Riccati ODE would be too narrow unless those assumptions were supplied.[1]
Likewise \(D=\mathbb R_+^m\times\mathbb R^n\) is the paper's main canonical setting, not a claim that the idea is meaningful nowhere else. Its §12 extends the transform definition to other state spaces. In a concrete model, one must name \(D\), the allowed starting states and whatever parameter constraints make the transition law valid there.[1]
Manages Complexity¶
The transform identity replaces an otherwise large family of transition distributions with two functions, \(\phi(t,u)\) and \(\psi(t,u)\), whose dependence on the initial state is fixed. In regular models, generator and Riccati characterizations provide a way to compute them. The simplification is genuine but conditional: an invalid boundary parameter, missing moment condition or unlicensed pricing assumption cannot be repaired merely by writing an affine formula.[1]
This separation makes a useful audit possible. First ask whether the process law is Markov and its characteristic transform affine in the initial state. Then ask whether the additional regularity for a Riccati method is established. Only afterward ask whether a specific observable—such as a discounted cash flow or a first-event probability—has an appropriate transform representation.[1][2]
Abstract Reasoning¶
Let \(P_t(x,dy)\) be the Markov transition kernel and \(f_u(y)=e^{i\langle u,y\rangle}\). The identity \(P_tf_u(x)=e^{\phi(t,u)+\langle\psi(t,u),x\rangle}\) constrains how the entire transition law changes with starting state. It does not specify a particular sample path. For a regular affine class, the semigroup's action on these exponential test functions closes into generalized Riccati equations; the closure is the source of tractability.[1]
For an OU diffusion, the solution can be written \(X_t=e^{-\kappa t}x+Z_t\) after absorbing its constant mean term and Brownian innovation into \(Z_t\), which is independent of the starting state \(x\). Thus \(\mathbb E_x[e^{iuX_t}]=\mathbb E[e^{iuZ_t}]e^{iu e^{-\kappa t}x}\): its initial-state exponent is affine. A nonnegative square-root/CBI realization gets the same abstract transform form by a different transition mechanism and admissibility conditions.[1]
Knowledge Transfer¶
The transferable test is not “does this model price a bond?” but “does the conditional law carry an exponential-affine signature across starting states?” That allows an OU transition law and a nonnegative branching-type diffusion to be classified together while keeping their path, boundary and application differences visible. A jump-diffusion from the frozen Wikipedia page enters only if it passes the same transform test; its jump mechanism is a subtype feature.[1][2]
Conversely, an apparently simple Markov model with nonaffine transition dependence should not be imported into this class merely because one local linearization or fitted regression is affine. The distinction preserves the mathematical promise of the label.[1]
Examples¶
Gaussian OU transition. The Markov family is a mean-reverting diffusion on \(\mathbb R\). The comparison frame fixes horizon \(t\) and Fourier argument \(u\) while varying initial state \(x\). The transform factors into an innovation term independent of \(x\) and \(e^{iu e^{-\kappa t}x}\), satisfying the required exponential-affine form. This is a continuous, jump-free case; its staged OU entry is a narrower process identity, not an alias of this class.[1]
Mapped back: all three defining roles are visible. No bond price, time-integral transform or Feller boundary is needed.
Nonnegative square-root/CBI diffusion. The Markov family evolves on \(\mathbb R_+\) under admissible coefficients. The comparison frame again varies nonnegative \(x\) at a fixed \(t,u\). The transform has \(e^{\phi(t,u)+\psi(t,u)x}\) form, though its \(\psi\) follows model-specific regular affine/Riccati equations rather than the OU formula. CIR is a familiar square-root member when its parameters define the intended nonnegative diffusion.[1][2]
Mapped back: the same three roles occur despite non-Gaussian state geometry. The seed's claim that a Feller strict-positivity inequality is constitutive of every affine process is rejected.
Negative boundary. A Markov process supplies a transition kernel, but suppose its conditional characteristic function cannot be represented by \(e^{\phi(t,u)+\langle\psi(t,u),x\rangle}\) over its declared state space. It satisfies the genus but lacks the affine differentia. A locally linear approximation at one \(x\) would not reverse that verdict.[1]
Structural Tensions¶
- Tractability versus flexibility. The fixed transform shape makes analysis feasible but can exclude dynamics a modeler might otherwise want. Diagnostic: Is the identity verified over the declared state space, or only a convenient local fit?[1]
- Transform definition versus generator method. Riccati machinery is powerful for regular classes, but its theorem has hypotheses. Diagnostic: Which regularity and admissibility results license a proposed generator or ODE characterization?[1]
- Common stochastic class versus application-specific output. OU, square-root and jump realizations share a transform signature, yet a bond or default calculation adds economic structure. Diagnostic: What rate, intensity, payoff and existence assumptions connect the state law to the claimed output?[1][2]
Structural–Framed Character¶
Affine Process is mixed-structural, leaning formal. Its exponential-affine transform condition is exact once a Markov state space and admissible transition family are fixed, but those modeling choices are not supplied by the name alone. Its evaluative weight is low: an affine process need not fit data well or be the best pricing model. It is partly human-practice-bound as a mathematical classification—someone specifies a state, horizon and transform—but a stochastic system can satisfy or fail the stipulated relation independently of anyone's preference. Its institutional origin is probability theory, later heavily used in mathematical finance; finance is not part of the definition. Its vocabulary travel extends from OU-type and branching processes to different state-space geometries, yet “affine” here means a conditional characteristic-transform form, not any linear-looking system. Import versus recognition requires that the same transform identity hold across initial states; calling an unrelated organizational trajectory “affine” would only import a label.
The portable skeleton is live Markov Process: the current state determines the transition law without a need for earlier history once the state is sufficiently specified. Affineness is a stricter condition on that law, not a new universal story about change. A valid coordinate choice or state augmentation must be declared before testing it. Its character: a formal stochastic-process subclass with real cross-model recognition but with its technical identity pinned to state-conditioned probability transforms.
Structural Core vs. Domain Accent¶
This decomposition tests whether the broader transferable lesson is already covered by a prime.
What is skeletal. A sufficiently specified present state determines future transition probabilities. That is live Markov Process, the strict parent here. It travels across physical, computational and biological models without requiring a particular transform formula. Affine Process narrows that broad conditional-state structure by demanding one state-wide functional form of the transition characteristic function.
What is domain-bound. The conditional transform must be expressible as an exponential with an exponent affine in the initial state, over an admissible collection of states, horizons and arguments. Remove that relation and the model can remain Markov while ceasing to be affine in this technical sense. OU-type real-valued motion, nonnegative branching or CIR-style diffusion, jump versus continuous paths, regular generator formulas, and particular Riccati implementations are alternative model families or qualified characterizations. Bond pricing, default modeling, volatility and Fourier inversion attach financial quantities and payoffs; they cannot be read backward as necessary roles of the process class. A fitted linear-looking drift or one selected initial state is not a substitute for the state-wide transform test.
Why this is not a prime. The general present-state screening relation has the cross-substrate reach of Markov Process. The additional exponential-affine transform is recognized across stochastic models but remains a mathematical-probability condition; using “affine” for a linear trend in another domain does not instantiate it. Promoting this entire technical subclass to a prime on the strength of Markov portability would transfer the parent's reach to a child whose distinctive vocabulary and test do not independently travel that far.
Instantiates / Related Primes¶
This entry is a kind of Markov Process.
Proposed strict subsumption: live Markov Process (Markov Process). The child preserves current-state stochastic transitions and adds a condition on their conditional transform. Live Stochastic Process (Stochastic Process) is a broader ancestor through Markov Process. Staged Ornstein–Uhlenbeck Process (Ornstein–Uhlenbeck Process) is a narrower example, not a parent or synonym. No catalog edge is applied outside this workspace.
Relationships to Other Abstractions¶
Current abstraction Affine Process Domain-specific
Parents (1) — more general patterns this builds on
-
Affine Process is a kind of Markov Process Prime
Every affine process is a Markov process with the additional exponential-affine transition-transform condition.Live Markov Process supplies a stochastic state-transition family whose current state screens off prior history. Affine Process preserves that structure and requires a distinctive conditional characteristic-function form across initial states. A Markov process need not satisfy that form, while without Markov transitions the affine-process definition here is not met.
Hierarchy paths (4) — routes to 4 parentless roots
- Affine Process → Markov Process → Stochastic Process
- Affine Process → Markov Process → State and State Transition → Phase Space
- Affine Process → Markov Process → Probability → Measure → Set and Membership
- Affine Process → Markov Process → Probability → Measure → Aggregation → Micro Macro Linkage
Neighborhood in Abstraction Space¶
Affine Process sits in a sparse region of the domain-specific corpus (77th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Unclustered & Miscellaneous (2551 abstractions)
Nearest neighbors
- Kolmogorov Equations for Continuous-Time Markov Chains — 0.85
- Control-Theoretic Orbit — 0.85
- Metropolis Algorithm — 0.82
- Differential equation — 0.82
- Differential Inclusion — 0.82
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
An affine jump diffusion is a subclass, not the entire transform-defined family. A state-affine drift is insufficient by itself. A financial affine-pricing formula is an application only when the stochastic process and extra payoff/discount assumptions warrant it. Regular-affine generator or Riccati theorems should not be universalized to the bare definition without their assumptions.[1][2]
References¶
[1] Darrell Duffie, Damir Filipović and Walter Schachermayer, “Affine Processes and Applications in Finance,” Annals of Applied Probability 13 (2003), 984–1053, author-hosted original full text directly checked: Definition 2.1/Eq. (2.2) PDF p. 6; Definitions 2.4–2.5 and Theorem 2.7 PDF pp. 6–8; Corollary 2.10 and irregular counterexample PDF pp. 8–9; generalized state-space Definition 12.1 PDF pp. 46–47; financial applications §§11, 13. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w ↩x ↩y
[2] Darrell Duffie, Jun Pan and Kenneth Singleton, “Transform Analysis and Asset Pricing for Affine Jump-Diffusions,” author-hosted 1999 draft of Econometrica 68 (2000), abstract and introduction PDF pp. 1–2 directly checked for the narrower affine-jump-diffusion model family and conditional valuation applications. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i