homogenization

One correction, many fields

A survey of fast random switching and its first correction, from option pricing to spectroscopy, gene expression and machine learning.

Many fields study a slow quantity driven by a fast hidden process, and they reach the same two results. The leading answer replaces the hidden process by its average. The first correction is an integrated correlation of the fluctuations, a Green–Kubo integral.

The table gives the name each field uses for the two results. Every reference below is on the bibliography page with a checked DOI.

The common object

A hidden state $y_t$ moves fast, with stationary law $\pi$. A slow quantity evolves under an operator that depends on the hidden state:

$$\mathcal L_y = \bar{\mathcal L} + \sum_k \phi_k(y)\,A_k .$$

The averaged model uses $\bar{\mathcal L}$. The first correction adds

$$\sum_{j,k} K_{jk}\,A_j A_k, \qquad K_{jk} = \int_0^\infty \operatorname{Cov}_\pi\big(\phi_j(y_0),\,\phi_k(y_\tau)\big)\,d\tau ,$$

as derived on the any-equation page. Fields differ in what the slow quantity is, what the operators $A_k$ are, and what can be observed.

fieldfast processslow quantityaveraged lawfirst correction
option pricingvolatility factor or regimeoption priceBlack–Scholes at effective volatilitysmile and skew terms
term structureregime of level and volatilitybond priceVasicek with averaged parametersthree Green–Kubo numbers
creditcommon frailty or regimejoint survivalindependent namesdefault dependence
turbulent transportsmall-scale velocityscalar or large-scale flowtransport by the mean floweddy diffusivity and eddy viscosity
statistical physicsmicroscopic fluctuationscurrentsno net transporttransport coefficients
spectroscopyhopping resonance frequencyline shapea sharp line at the mean frequencymotional-narrowing width
chromatography, dispersionphase or streamlinesolute banddrift at mean speedeffective dispersion
gene expressionpromoter statemolecule countsPoisson at mean rateexcess noise
population biologyenvironmentpopulation sizegrowth at mean rateautocorrelation correction to growth
choice and racesshared driverwho arrives firstLuce's choice axiomcorrection to softmax
learningdata streamparametersgradient flow on the mean lossfluctuation and ordering terms

Mathematical foundations

The averaging principle for equations with fast random coefficients goes back to Khasminskii (1966). He showed that the fluctuations about the average converge to a diffusion whose coefficients are integrated correlations. Papanicolaou and coauthors built the general theory in the 1970s:

The same structure appears as random evolutions: operators switched by a Markov chain, whose expectations solve coupled systems of equations (Griego and Hersh 1969; Hersh 1974; Pinsky 1991). Kurtz gave the operator-semigroup limit theorem for generators of the form $Q/\varepsilon + \mathcal L$ (Kurtz 1973; Ethier and Kurtz 1986). Every pricing equation in the examples is a random evolution.

The long-run covariance behind the Green–Kubo correction is the subject of the central limit theorem of Kipnis and Varadhan (1986), proved through the resolvent for additive functionals of reversible Markov processes. The finite chains here need not be reversible, and for them the group inverse gives the same variance directly.

Physicists reached the same correction through cumulant expansions. The second cumulant of a linear equation with random coefficients is a time-ordered product of the random operators, weighted by their correlation (Kubo 1962; van Kampen 1974). That is the first-order rule with non-commuting operators, including the commutator term.

Full asymptotic series for fast Markov chains, with outer expansions and initial layers, are developed in:

Control theory reached aggregation of fast states independently (Phillips and Kokotovic 1981).

For diffusions, the Poisson equation that plays the role of $Q w = f$ is solved by Pardoux and Veretennikov (2001). Periodic homogenization, the spatial version of the same idea, is the subject of Bensoussan, Lions and Papanicolaou (1978). Three textbooks cover the methods: Freidlin and Wentzell (1984), Skorokhod, Hoppensteadt and Salehi (2002) and Pavliotis and Stuart (2008).

The diffusion approximation was developed at length for waves in randomly layered media (Fouque, Garnier, Papanicolaou and Sølna 2007). Fast Markov modulation of diffusions has recent limit theorems of its own (Huang, Mandjes and Spreij 2014; Huang et al. 2016).

Spatial and nonlinear homogenization

Homogenization of coefficients that oscillate in space is a separate theory, and it answers three different questions: what the effective coefficients are, in what sense solutions or energies converge, and how large the fluctuations and the correctors are. None of it involves a state that switches in time.

Finance: fast mean-reverting volatility

Fouque, Papanicolaou and Sircar brought the diffusion approximation to option pricing. Volatility is driven by a fast mean-reverting factor. The price is Black–Scholes at an effective volatility plus a correction (Sircar and Papanicolaou 1999; Fouque, Papanicolaou and Sircar 2000; their 2000 book).

In that setting the first correction is of order $\sqrt\varepsilon$ and needs correlation between the asset and the factor. It produces the skew and is summarized by a few group parameters that can be fitted to implied volatilities. The fast-factor page computes that series to all orders.

The program was extended in several directions:

The book Fouque, Papanicolaou, Sircar and Sølna (2011) collects equity, interest rate and credit.

The uncorrelated case goes back to the mixing formula of Hull and White (1987): average Black–Scholes over integrated variance. To second order in the variance of integrated variance, that formula is the smile on the any-equation page. Sircar and Papanicolaou (1999) derive this first-order smile for a fast ergodic driver whose integrated fluctuations obey a central limit theorem, including a finite-state Markov chain.

A nontrivial finite-chain feature has a positive Green–Kubo coefficient. More general drivers can behave differently: long-range dependence makes the correction larger, while zero-Green–Kubo antipersistence makes it smaller than order $\varepsilon$.

For a stationary scalar diffusion on an interval or the line, the usual zero-current boundary conditions imply reversibility and a symmetric Green–Kubo matrix. This is a boundary/current statement, not a consequence of dimension alone: every scalar periodic diffusion splits into a reversible part and a skew stationary-current part, with drifted Brownian motion as the closed-form example. A finite-state chain or periodic diffusion that cycles in one direction can therefore move the Heston smile when two non-commuting coefficients fluctuate together (cycle direction in the smile).

Finance: regime switching

For bond prices and survival probabilities, the expansion in the switching rate goes back to Cotton (2001). Its second half treats a Vasicek short rate or hazard rate with mean level and volatility switched by a two-state chain. It reduces the price to a linear system of ordinary differential equations and expands it for fast switching.

The regime-switching pages develop it to all orders, and the engine extends it to any finite chain and any exponential-affine model.

For option prices, Yin (2009) gives a full series in the switching rate with an averaged Black–Scholes leading term, and Basu and Ghosh (2009) study the fast and slow limits.

Regime-switching models were common in their own right, mostly without any fast-switching asymptotics. Regime switching entered econometrics with Hamilton (1989). Option pricing papers include:

Regime-switching term-structure models include:

Finance: credit

For credit, the fast-switching expansion also goes back to Cotton (2001). There the hazard rate is an Ornstein–Uhlenbeck process whose mean level and volatility are switched by a two-state chain, and the survival probability is expanded in the switching rate. The survival pages carry it to all orders, and the two-name credit example extends it to joint survival.

Reduced-form credit models treat default as the first event of a Cox process (Cox 1955; Lando 1998; Duffie and Singleton 1999). They are priced with affine transforms (Duffie, Pan and Singleton 2000).

Defaults cluster beyond observable covariates. Duffie, Eckner, Horel and Saita (2009) model the clustering with a common latent frailty.

With a fast frailty, the any-equation rule gives the joint survival of every set of names from one Green–Kubo matrix. With constant hazards this fixed matrix has rank at most one less than the number of regimes.

Positive CIR intensities with heterogeneous maturity loadings can produce a higher-rank integrated correction without violating that fixed-parameter bound. Which name defaults first is the race studied below: its cause probabilities expose the antisymmetric part, although several weighted races may be needed to identify that part completely.

Fluid dynamics and turbulence

Turbulent transport was a main driver of the theory. Taylor (1922) showed that a particle carried by a turbulent velocity spreads, at long times, with a diffusivity equal to the time integral of its velocity autocorrelation. That is the first Green–Kubo formula, three decades before Green and Kubo.

The probabilistic limit theorems grew up alongside. The martingale approach of Papanicolaou, Stroock and Varadhan (1977) appeared in the proceedings of a turbulence conference. Kraichnan (1968) introduced the random velocity, delta-correlated in time, in which the eddy diffusivity is exact.

For a passive scalar in a flow with small-scale structure, homogenization gives an effective diffusivity from a cell problem:

For the Navier–Stokes equations, multiscale expansion around a small-scale flow gives an eddy viscosity when the flow is parity invariant (Dubrulle and Frisch 1991). Without parity invariance it gives a large-scale instability already at first order (Frisch, She and Sulem 1987). The book of Frisch (1995) places these results in turbulence theory.

The nonlinearity is the main difference from the linear pricing equations. In the Navier–Stokes equations the fast modes are generated by the flow itself, and closing the equations for the slow modes is the closure problem.

Two recent works on three-dimensional Navier–Stokes are sometimes read alongside averaging, and neither is the fast switching of an exogenous state. Tao (2016) replaces the bilinear term by an averaged operator, built from rotations and Fourier multipliers, that keeps the energy identity, and proves finite-time blowup for that modified equation; the result does not transfer to the original equations.

A preprint of September 2026 (OpenAI 2026) claims finite-time blowup for the actual forced equations: a smooth, compactly supported force and zero initial velocity, with bounded energy and unbounded velocity, which are the forced alternatives of the Clay statement. As of September 2026 the Clay Mathematics Institute describes the problem as apparently settled and its evaluation as ongoing.

In both constructions the oscillations are produced by the flow and their Reynolds stress $\langle w \otimes w\rangle$ enters at leading order, so replacing the velocity by its mean drops a leading term. The $1/\lambda$ Green–Kubo correction for an exogenous chain does not apply to them without a separate scaling theorem.

Climate models apply the same elimination of fast modes. Hasselmann (1976) treated weather as noise driving slow climate variables. Majda, Timofeyev and Vanden-Eijnden (2001) derived the reduced stochastic equations systematically by the diffusion approximation.

Statistical physics

Transport coefficients are integrated equilibrium correlations. This is the Green–Kubo formula (Green 1954; Kubo 1957), and the matrix $K$ is named after it. Onsager showed that transport coefficients are symmetric under microscopic reversibility (Onsager 1931).

For a fast chain the same argument makes $K$ symmetric when the chain satisfies detailed balance. The antisymmetric part is therefore a measure of irreversibility.

Kinetics uses it to detect a hidden driver that is out of equilibrium, from the order of arrivals (Cotton, working draft). The any-equation page shows it shifting an effective variance.

Spectroscopy and relaxation

A nuclear spin whose resonance frequency hops between values gives a spectral line that narrows as the hopping speeds up. Anderson (1954) and Kubo (1954) explained this motional narrowing. The width of the narrowed line is the Green–Kubo integral of the frequency fluctuation.

Mathematically it is the linear system of the regime-switching examples with an imaginary forcing $i\omega(y)$. The two-state eigenvalue on the yield-curve page is Anderson's two-site formula.

Relaxation rates in magnetic resonance are spectral densities of fluctuating fields (Bloembergen, Purcell and Pound 1948). Redfield theory treats spin operators that do not commute (Redfield 1957). The rates come from the real parts of spectral densities, and the frequency shifts from their imaginary parts.

Blume (1968) extended the stochastic line-shape model to fluctuating Hamiltonians that do not commute at different times.

Transport, chemistry and motors

Solute in a tube spreads faster than molecular diffusion alone, because different streamlines move at different speeds (Taylor 1953; Aris 1956). In chromatography a molecule alternates between moving and stuck phases (Giddings and Eyring 1955).

The band moves at the averaged speed, and it spreads with the Green–Kubo integral of the velocity. This is the averaged law and first correction for a switched drift.

Reaction rates that fluctuate in time are called dynamical disorder (Zwanzig 1990). Reactions gated on and off by a random switch were studied by Szabo, Shoup, Northrup and McCammon (1982). A boundary that switches between absorbing and reflecting homogenizes to a Robin condition (Lawley and Keener 2015).

Molecular motors switch between potentials under a chemical cycle that breaks detailed balance (Jülicher, Ajdari and Prost 1997; Reimann 2002). They are the physical case where the antisymmetric part of $K$ drives the effect.

Biology

Bressloff (2017) reviews stochastic switching across biology.

In gene expression, a promoter switching between states modulates the rate of synthesis. This is the telegraph model (Peccoud and Ycart 1995). It is observed as bursting in single-cell counts (Raj et al. 2006).

The excess noise over Poisson is the Green–Kubo term for the switched rate, as on the Poisson counts page. At finite horizons the exact excess includes a boundary correction; more generally, count factorial cumulants equal the ordinary cumulants of the integrated stochastic rate, so higher-order upstream fluctuations can be separated from Poisson shot noise.

For multiple conditionally independent count streams the identity is coordinatewise; common events instead leave shared-shot-noise terms given by an exact all-order partition formula, including arbitrary integer-valued event marks through their falling-factorial moments. Paulsson (2004) analyses how time-averaging of fluctuating upstream signals controls noise in gene networks.

A population growing in a random environment grows in expectation at the averaged rate plus the integrated covariance of the rate. Lewontin and Cohen (1969) showed that the expected and typical growth rates differ.

Tuljapurkar (1982) gave the small-noise formula for structured populations in correlated environments, with a monograph in 1990. For structured populations the projection matrices do not commute, so the commutator term applies.

Kussell and Leibler (2005) connect growth in a switching environment to information.

Bacteria that run and tumble spread with an effective diffusion given by their integrated velocity correlation (Othmer, Dunbar and Alt 1988; Schnitzer 1993). Ion channels observed only as open or closed are aggregated Markov chains (Fredkin and Rice 1986). They raise the same question as the yield-curve page: what can be identified about a hidden chain from coarse observations.

Queues and point processes

Arrivals driven by a hidden chain form a Markov-modulated Poisson process (Fischer and Meier-Hellstern 1993), with marked versions in He and Neuts (1998). Their long-run count variance exceeds the Poisson variance by twice the Green–Kubo integral of the rate, and that excess drives queue lengths in heavy traffic (Asmussen 2003).

At finite horizons the exact excess is the variance of the cumulative intensity, including its initial-condition boundary term. A request sent to several servers, with the first reply winning, is a race under a shared driver (Dean and Barroso 2013).

Lumped chains and non-exponential durations

A regime whose duration is not exponential is a group of hidden states observed as one. Split each regime into stages with constant hazards and the stage chain is Markov, while the regime is a function of it.

When grouping the states of a chain gives another Markov chain is an old question (Burke and Rosenblatt 1958). The answer is the lumpability condition of Kemeny and Snell (1960): every state of a group must have the same total rate into each other group. The interesting regime models fail it, since that is what lets a regime's age matter.

Gilbert (1959) asked which hidden chains can be told apart from their grouped output. Heller (1965) characterized the processes that arise this way.

Durations built from exponential stages are the phase-type distributions of Neuts (1981). They can approximate any distribution on the positive axis, so a constant-hazard stage chain covers any regime duration to any accuracy.

Ion channels are the best-studied case. A channel is seen only as open or shut, while several hidden conformations set the dwell times (Colquhoun and Hawkes 1981). The structure of bursts and clusters of openings shows how the hidden states connect (Colquhoun and Hawkes 1982; Fredkin and Rice 1986).

For fast switching the observable is constant on each group, so the first-order correction stays a matrix over the regimes, however many stages there are. For two regimes the Green–Kubo number depends on the durations only through their means $m_i$ and standard deviations $s_i$:

$$K = (\theta_1 - \theta_2)^2\,\frac{m_2^2 s_1^2 + m_1^2 s_2^2}{2\,(m_1 + m_2)^3} .$$

The regular durations page works this out for a yield curve. There the age of the current regime enters the price through the start term.

Choice and competing arrivals

When channels fire as Poisson processes with rates driven by a shared fast driver, the probability that a channel fires first tends to Luce's choice axiom in the mean rates (Luce 1959). The kinetics site derives the first correction, governed by the same matrix $K$ (Cotton, working draft). It shows three further results:

The number of observed channels and the number of hidden states play different roles. For two hazards $h_1,h_2$ and a stationary hidden chain with fast generator $Q_m=mQ$, the exact first-arrival probability is

$$p_1(m)=\pi^\top\big(\operatorname{diag}(h_1+h_2)-Q_m\big)^{-1}h_1.$$

Writing $\bar h=\bar h_1+\bar h_2$ and computing the directed Green–Kubo matrix $K$ for $Q_m$, its first-order expansion is

$$p_1(m)=\frac{\bar h_1}{\bar h} +\frac1{\bar h}\left\{\frac{\bar h_1}{\bar h}\sum_{j,k=1}^2K_{jk}-K_{11}-K_{21}\right\} +O(m^{-2}).$$

To see this, solve $(\operatorname{diag}(h_1+h_2)-mQ)u=h_1$ and insert $u=u_0\mathbf1+m^{-1}u_1+\cdots$. The first solvability condition gives $u_0=\bar h_1/\bar h$; the next Poisson equation, followed by its solvability condition, gives the displayed Green–Kubo coefficient.

Every order, with a finite-rate bound

The same calculation has an exact centered resolvent form. Put $d=h_1+h_2$, $R=Q^{\#}$,

$$c_0=\frac{\pi h_1}{\pi d},\qquad \ell^\top=\frac{\pi\operatorname{diag}(d)}{\pi d},\qquad B=I-\mathbf1\ell^\top,$$ $$a=R(dc_0-h_1),\qquad M=R\operatorname{diag}(d)B.$$

Splitting $u$ into its stationary scalar and centered parts gives the exact identity

$$p_1(m)=c_0-\frac1m\ell^\top\left(I-\frac{M}{m}\right)^{-1}a.$$

Consequently, whenever $m>\rho(M)$,

$$p_1(m)=c_0-\sum_{k\ge1}\frac{\ell^\top M^{k-1}a}{m^k}.$$

This supplies every inverse-speed coefficient without repeatedly solving Poisson equations. More importantly, truncation after order $N$ has the exact remainder

$$\mathcal R_N=-\frac1{m^{N+1}}\ell^\top M^N \left(I-\frac{M}{m}\right)^{-1}a.$$

The directly checkable sufficient condition $m>\lVert M\rVert_2$ therefore gives the nonasymptotic bound

$$|\mathcal R_N|\le \frac{\lVert\ell\rVert_2\lVert a\rVert_2}{m} \frac{(\lVert M\rVert_2/m)^N}{1-\lVert M\rVert_2/m}.$$

The norm condition is sufficient rather than necessary; the sharp power-series radius is controlled by $\rho(M)$. For the three-state race below, $\rho(M)=0.01613437$ and $\lVert M\rVert_2=0.02251066$.

Already at speed $m=0.1$, the fourth-order error is $2.20\times10^{-5}$ against the rigorous bound $2.53\times10^{-4}$; at $m=10$ it is below $10^{-14}$. The first coefficient is algebraically the Green–Kubo formula above.

Exactly when two channels see the arrow of time

Let $Q^\dagger=\operatorname{diag}(\pi)^{-1}Q^\top\operatorname{diag}(\pi)$ be the stationary time reversal (Kelly 1976). Its directed Green–Kubo matrix is $K^\top$. Hence, for two hazards,

$$p_{1,Q_m}-p_{1,Q_m^\dagger} =\frac{K_{12}-K_{21}}{\bar h}+O(m^{-2}).$$

Here $K$ is computed for the already-scaled generator $Q_m=mQ$, so its entries are $O(m^{-1})$. Thus a fixed pair sees the arrow of time at first order if and only if $K_{12}\ne K_{21}$. Merely having three hidden states and violating detailed balance is not sufficient for an arbitrary pair of hazards.

The existential statement is exact: an irreducible finite-state chain is reversible if and only if $K(f,g)=K(g,f)$ for every pair of state observables. On the centered subspace, $Q^\#$ is the inverse of $Q$; therefore symmetry of every Green–Kubo pairing is equivalent to self-adjointness of $Q^\#$, hence of $Q$, in the stationary inner product.

If $Q$ is not reversible, choose $f,g$ with a nonzero antisymmetric pairing and add sufficiently large constants to make both positive. Two positive arrival channels then witness the irreversibility.

There is also an exact blind family. Put $H=h_1+h_2$ and $A=\operatorname{diag}(H)-mQ$. If $h_1=\alpha H+\beta\mathbf1$ — equivalently, the two hazards are affine functions of one another — then

$$p_{1,Q_m}=\langle\mathbf1,A^{-1}h_1\rangle_\pi =\langle A^{-1}\mathbf1,h_1\rangle_\pi=p_{1,Q_m^\dagger}$$

for every switching speed, not just asymptotically. The two equalities use $A\mathbf1=H$ and $A^\dagger\mathbf1=H$. The certificate checks this exact blindness for a nonreversible three-state chain with a nonuniform stationary law.

For a visible example, a three-state cycle with rates $2$ clockwise and $0.1$ counterclockwise, $h_1=(0.07,0.01,0.01)$ and $h_2=(0.01,0.07,0.01)$ has a nonzero pairing. The certificate gives $p_1(10)=0.5001500361$; reversing the cycle gives $0.4998499639$. The first-order and reversal-gap errors both fall by a factor of four when $m$ doubles, confirming the $O(m^{-2})$ remainder (certificate).

The horse-race model of choice and response times is its psychological ancestor (Marley and Colonius 1992).

Learning and stochastic approximation

Algorithms fed by a Markovian data stream are analysed by averaging. The parameters follow the ordinary differential equation of the mean update, and the fluctuations about it have integrated-correlation covariances (Benveniste, Métivier and Priouret 1990; Kushner and Yin 2003). Convergence rates for data from an ergodic process depend on its mixing time (Duchi, Agarwal, Johansson and Jordan 2012).

The order in which data is visited matters. Reshuffling beats independent sampling (Gürbüzbalaban, Ozdaglar and Parrilo 2021).

Per-example updates act like switched operators whose Hessians do not commute. The rule then suggests that a structured data order would enter through the antisymmetric part of the correlation matrix. This has not been checked.

Results beyond the classical theory

The averaged law, the Green–Kubo first correction, and its non-commuting form are classical. The following results do not appear in the literature above.

Open directions