Affine Model: Structures and Transformations

• The affine model is a family of structures organized by affine combinations, transformations, or exponential-affine transforms, used across various disciplines.
• In finance and stochastic processes, affine models enable tractable exponential-affine transforms governed by Riccati-type equations, facilitating robust term-structure modeling.
• In regression and continuous logic, affine models provide linear-plus-offset predictors and formula algebras, enhancing computational efficiency and domain transfer.

An affine model is a domain-dependent construction in which the governing object is organized by affine combinations, affine transformations, or exponential-affine transforms. In stochastic-process and mathematical-finance literature, the defining property is typically exponential-affine dependence of characteristic or Laplace transforms on the current state (Keller-Ressel et al., 2018). In regression and transfer learning, it denotes linear-plus-offset predictors or affine transformations of source models and features (Minami et al., 2022). In continuous logic and information geometry, it refers to affine operations, affine bundles, and affine transport structures rather than multiplicative or lattice connectives (Bagheri, 2024, Pistone, 2022). This suggests that “affine model” is not a single formalism but a family of structures whose common feature is affine organization of state, transform, or representation.

1. General concept and scope

Across the cited literature, three recurrent meanings appear. First, an affine stochastic model is one whose key transform has the form

$$\mathbb{E}\!\left[e^{u\cdot X_T}\mid X_t=x\right]=\exp(\phi(\tau;u)+\psi(\tau;u)\cdot x),$$

with $\tau=T-t$, and $\phi,\psi$ governed by Riccati-type equations (Keller-Ressel et al., 2018). Second, in supervised regression an affine model is explicitly “linear plus offset,” with representative forms

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$

(Minami et al., 2022). Third, in affine continuous logic, the affine fragment is obtained by removing $\wedge,\vee$ and retaining affine operations together with $\sup/\inf$ quantifiers (Bagheri, 2024).

Domain	Affine object	Characteristic form
Stochastic processes	Conditional transform	$\exp(\phi+\psi\cdot x)$
Transfer learning	Target predictor	$g_1(f_s(x))+g_2(f_s(x))g_3(x)$
Continuous logic	Formula algebra	affine operations and $\sup/\inf$

A common misconception is that “affine” always means merely “linear plus intercept.” That is accurate in regression, but in finance the term is attached to transform formulas, not necessarily to linear state dynamics (Keller-Ressel et al., 2018). Conversely, in logic the term refers to the algebra of formulas and type spaces, not to Euclidean affine geometry (Bagheri, 2024).

2. Affine stochastic processes and term-structure models

In stochastic processes and mathematical finance, the central notion is exponential-affine transformability. Classical affine stochastic volatility models, including Heston-type systems, admit transforms of the form

$$\mathbb{E}[e^{uL_T}\mid \mathcal{F}_t] = \exp\big(uL_t+\phi(T-t;u)+\psi(T-t;u)V_t\big),$$

with $\tau=T-t$0 and $\tau=T-t$1 governed by Riccati ODEs (Keller-Ressel et al., 2018). In rough and Volterra settings, the Riccati ODE is replaced by a Riccati–Volterra or fractional Riccati equation, while the transform remains exponential-affine in the log-price and forward variance curve (Keller-Ressel et al., 2018).

Time-inhomogeneous affine processes generalize the homogeneous case by allowing $\tau=T-t$2 and $\tau=T-t$3 to depend on both dates. Under stochastic continuity they admit a càdlàg modification and a strong Markov property, but, unlike the homogeneous setting, regularity and the semimartingale property are not automatic (Waldenberger, 2015). When a time-inhomogeneous affine process is a semimartingale of finite variation type, the affine parameters satisfy generalized Riccati integral equations

$\tau=T-t$4

with deterministic time change $\tau=T-t$5 (Waldenberger, 2015). This provides a tractable extension of the classical affine-transform framework to nonstationary settings such as affine LIBOR and affine inflation market models.

Several specialized affine process families appear in the literature. A two-dimensional affine model with an $\tau=T-t$6-root process in the first coordinate has a unique stationary distribution for $\tau=T-t$7, and in the diffusion case $\tau=T-t$8 it is also exponentially ergodic (Barczy et al., 2013). Under parameter uncertainty, one-dimensional non-linear affine processes replace a single generator by the supremum over affine generators indexed by a compact parameter set, yielding a variational Kolmogorov PDE and robust term-structure equations (Fadina et al., 2018). In interest-rate modeling, an affine extension of the Linear Gaussian term structure Model equips the Gaussian factors with an affine covariance process on positive semidefinite matrices, preserving exponential-affine bond pricing while generating caplet and swaption smiles (Ahdida et al., 2014).

These uses share a precise structural point: tractability comes from affine transform formulas and Riccati dynamics, not from linear sample-path evolution. That distinction is central to the finance literature (Keller-Ressel et al., 2018, Waldenberger, 2015).

3. Affine forward variance and affine forward intensity

The paper “Affine forward variance models” develops a forward-variance formulation in which the affine property is imposed on the conditional cumulant generating function rather than directly on a finite-dimensional Markov state (Gatheral et al., 2018). Let the forward variance curve be

$\tau=T-t$9

A forward variance model $\phi,\psi$0 has an affine cumulant generating function if

$\phi,\psi$1

The main characterization is that, under Assumption 2.1, the model is affine if and only if

$\phi,\psi$2

for a deterministic, decreasing $\phi,\psi$3-kernel $\phi,\psi$4 (Gatheral et al., 2018). The function $\phi,\psi$5 is then the unique global continuous solution of the convolution Riccati equation

$\phi,\psi$6

This replaces the finite-dimensional Riccati ODE by a Volterra-type object. The spot variance satisfies the affine Volterra representation

$\phi,\psi$7

Both the conventional Heston model and the rough Heston model are special cases. For Heston, the forward kernel is exponential, $\phi,\psi$8, and the usual Riccati ODE is recovered (Gatheral et al., 2018). For rough Heston,

$\phi,\psi$9

and the Riccati equation becomes fractional, involving the Riemann–Liouville derivative $$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$0 (Gatheral et al., 2018). This places Markovian and rough regimes in a single affine transform framework.

The same paper introduces affine forward order flow intensity models, or AFI models, which are jump-driven analogues of AFV models. Their cumulant generating function satisfies a generalized convolution Riccati equation with jump driver

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$1

(Gatheral et al., 2018). AFI models include Hawkes-type systems, and exponential or Mittag–Leffler Hawkes kernels produce, respectively, the Heston and rough Heston forward kernels. Under high-frequency scaling, AFI converges in distribution to AFV, with effective correlation

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$2

(Gatheral et al., 2018). A plausible implication is that the affine forward-variance formalism provides a direct bridge between microstructural order-flow models and macroscopic stochastic volatility.

4. Affine models in regression and transfer learning

In supervised regression, the affine model transfer framework derives an optimal affine transformation law under expected squared loss (Minami et al., 2022). The target-domain regression model is

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$3

with source features $$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$4 and transformation functions $$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$5 and $$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$6. Under differentiability, invertibility, and consistency assumptions, the paper proves that

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$7

and therefore the optimal target predictor has the form

$$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$8

(Minami et al., 2022). Here $$f(x)=\alpha g(x)+\beta,\qquad \theta_T=A\theta_S+b,\qquad f_T(x)=w^\top\phi(x)+b$$9 and $\wedge,\vee$0 encode inter-domain commonality, while $\wedge,\vee$1 captures domain-specific factors.

The empirical objective with RKHS regularization is

$\wedge,\vee$2

optimized by block relaxation over $\wedge,\vee$3 (Minami et al., 2022). The paper gives a generalization bound whose dominant term improves when the source-only risk $\wedge,\vee$4 is small, and an excess-risk rate

$\wedge,\vee$5

under eigenvalue decay assumptions (Minami et al., 2022).

This framework subsumes direct learning, offset transfer, scale transfer, frozen-feature affine heads, and certain Bayesian transfer schemes (Minami et al., 2022). It also separates commonality from target-specific effects in a way that mitigates negative transfer. On SARCOS, the method avoids negative transfer when source-target relation is weak and improves RMSE when it is strong; on SciRepEval it improves or matches frozen-feature baselines across BERT, SciBERT, T5, and GPT-3 embeddings (Minami et al., 2022). The paper is explicit, however, that the full affine optimality result depends on squared loss; outside that setting, $\wedge,\vee$6 persists more broadly, but the affine form is not generally optimal (Minami et al., 2022). That caveat corrects a common overgeneralization.

5. Affine logic, arithmetic, and computation

Affine continuous logic is the fragment of continuous logic obtained by avoiding $\wedge,\vee$7 and retaining affine operations together with $\wedge,\vee$8 quantifiers (Bagheri, 2024). Its ultraproduct analogue is the ultramean construction, in which ultrafilters are replaced by maximal finitely additive probability measures. For a family $\wedge,\vee$9 indexed by $\sup/\inf$0 and an ultracharge $\sup/\inf$1,

$\sup/\inf$2

and the affine Łoś theorem states that

$\sup/\inf$3

for every affine formula $\sup/\inf$4 (Bagheri, 2024). Type spaces become compact convex sets, extreme types play a central role, and compact structures with at least two elements have proper elementary extensions. This makes affine continuous logic strictly weaker than full continuous logic in the sense of elementary equivalence (Bagheri, 2024).

The affine part of Peano arithmetic, denoted AA, transports this perspective into arithmetic (Bagheri, 25 Aug 2025). Its language is

$\sup/\inf$5

with lattice operations replacing primitive order and with a nontrivial metric (Bagheri, 25 Aug 2025). Classical PA-models are exactly the linearly ordered, extremal AA-models, while nonclassical affine models arise via ultrameans and need not be linearly ordered. The paper proves affine versions of the least number principle, Euclidean division, Bézout’s theorem, the Chinese remainder theorem, prime existence, coding, factorial and exponentiation, and overspill/underspill (Bagheri, 25 Aug 2025). This suggests that affine logic can preserve substantial arithmetic content while enlarging the model-theoretic spectrum.

Affine computation gives yet another meaning: state vectors may have negative entries but must sum to $\sup/\inf$6, evolution is by matrices whose columns sum to $\sup/\inf$7, and measurement uses $\sup/\inf$8 weighting

$\sup/\inf$9

(Díaz-Caro et al., 2016). Affine finite automata are more powerful than PFAs and QFAs in bounded and unbounded error modes, and satisfy

$\exp(\phi+\psi\cdot x)$0

in the nondeterministic mode (Díaz-Caro et al., 2016). Here “affine” refers to barycentric-preserving linear dynamics rather than to geometry or statistical manifolds.

6. Geometry, physics, and structured applied models

Several specialized literatures use “affine model” for constrained geometric dynamics. In video coding, the efficient four-parameter affine motion model constrains a six-parameter affine transform to rotation, uniform zoom, and translation, with motion field

$\exp(\phi+\psi\cdot x)$1

(Li et al., 2017). Implemented in an HEVC-based codec, the full framework achieves on average $\exp(\phi+\psi\cdot x)$2 and $\exp(\phi+\psi\cdot x)$3 bits saving for random access and low delay configurations, respectively, on typical sequences rich in rotation or zooming motion (Li et al., 2017).

In vision, the affine Gaussian derivative model replaces isotropic Gaussian kernels by affine Gaussian kernels with covariance $\exp(\phi+\psi\cdot x)$4, preserving affine covariance under

$\exp(\phi+\psi\cdot x)$5

(Lindeberg, 2017). Discrete implementations are derived from a semi-discrete affine diffusion equation and from $\exp(\phi+\psi\cdot x)$6 kernels, with non-enhancement of local extrema preserved under positivity conditions. The eccentricity bound

$\exp(\phi+\psi\cdot x)$7

marks the range where non-negative discrete kernels exist for all orientations (Lindeberg, 2017).

In gravity, a polynomial purely affine model takes the affine connection, rather than a metric, as the fundamental field (Castillo-Felisola et al., 2014, Castillo-Felisola et al., 2015). In the torsion-free equi-affine sector, the effective equation reduces to

$\exp(\phi+\psi\cdot x)$8

so general Einstein manifolds, with or without cosmological constant, solve the theory (Castillo-Felisola et al., 2015). The nonrelativistic limit yields Newtonian gravity (Castillo-Felisola et al., 2014), and a Birkhoff-like theorem holds in the analyzed torsion-free sector (Castillo-Felisola et al., 2015). A different affine-group formulation of gravity coupled to the Standard Model rewrites the Hamiltonian constraint in an affine algebra involving a positive volume operator, producing a fundamental uncertainty relation tied to a non-vanishing cosmological constant (Chou et al., 2013).

In astrophysical fluid dynamics, the affine disc model treats each fluid column as undergoing a time-dependent affine transformation in three dimensions,

$\exp(\phi+\psi\cdot x)$9

thereby extending two-dimensional thin-disc hydrodynamics to warped, eccentric, and vertically breathing discs while preserving conservation of energy and potential vorticity (Ogilvie, 2018).

In cluster algebra theory, the affine almost positive roots model defines the set $g_1(f_s(x))+g_2(f_s(x))g_3(x)$0 of almost positive Schur roots, a compatibility degree $g_1(f_s(x))+g_2(f_s(x))g_3(x)$1, and the complete fan $g_1(f_s(x))+g_2(f_s(x))g_3(x)$2 (Reading et al., 2017). Every vector has a unique cluster expansion, and a piecewise-linear map $g_1(f_s(x))+g_2(f_s(x))g_3(x)$3 identifies the real subfan with the $g_1(f_s(x))+g_2(f_s(x))g_3(x)$4-vector fan of the associated acyclic cluster algebra (Reading et al., 2017).

Finally, in infinite-dimensional information geometry, an affine statistical bundle over a qualified set of probability densities uses exponential and mixture charts

$g_1(f_s(x))+g_2(f_s(x))g_3(x)$5

with fibers modeled on Gaussian Orlicz–Sobolev spaces and with Fisher’s score furnishing the statistically natural displacement (Pistone, 2022). The resulting structure is dually flat in the non-parametric sense.

Taken together, these usages show that “affine model” is a strongly polysemous technical term. In each case, however, the defining move is structurally similar: replace a more general object by an affine transform law, an affine chart system, or an exponential-affine transform, and use that structure to obtain tractability, invariance, or a complete combinatorial description.