The Ornstein-Uhlenbeck Process

The Ornstein-Uhlenbeck process was born in 1930 from a very concrete difficulty: the Wiener process, the idealized model of Brownian displacement, has nowhere-differentiable paths, so the velocity of a Brownian particle is mathematically undefined. Ornstein and Uhlenbeck fixed this by adding what physics demands anyway — linear friction. The velocity then becomes a smooth, mean-reverting Gaussian process with a finite variance pinned by temperature. The result is the natural bridge between the singular Wiener process and physically realistic, exponentially correlated fluctuations, and it has become the single most important exactly solvable model in the theory of stochastic processes. Mathematically the process is defined by the linear stochastic differential equation $\mathrm dv=-\gamma v\,\mathrm dt+\sqrt{2D_v}\,\mathrm dW$ — a Langevin equation, recalled by name, with a restoring drift. Linearity makes it exactly solvable: the solution is an explicit Gaussian functional of the driving noise, so every moment and correlation function follows in closed form. The drift $-\gamma v$ pulls the velocity back toward zero (mean reversion) while the noise spreads it out; their balance produces a stationary state with finite variance $k_BT/m$ — a microcosm of the fluctuation-dissipation balance underlying all of equilibrium statistical mechanics. The process earns its central place through a theorem of Doob: the Ornstein-Uhlenbeck process is the unique stationary Gaussian Markov process with continuous paths. That uniqueness forces its autocorrelation to be a single decaying exponential, $\langle v(t)v(0)\rangle=(k_BT/m)e^{-\gamma|t|}$ , whose Fourier transform is a Lorentzian power spectrum — the signature of a single relaxation rate with no memory. Because it is at once physically motivated (the velocity of a Brownian particle), mathematically canonical (Doob's uniqueness), and exactly solvable, the Ornstein-Uhlenbeck process is the standard testbed for the whole machinery of the field: Itô calculus, the Fokker-Planck equation, the fluctuation-dissipation theorem, and linear response all become transparent here. Beyond physics, the same equation is the Vasicek interest-rate model and the leaky integrate-and-fire neuron — wherever one degree of freedom relaxes linearly while buffeted by fast random kicks.

Key takeaways

The Ornstein-Uhlenbeck (OU) process is the unique stationary, Gaussian, Markov process with continuous paths; it solves the linear Langevin equation $\mathrm dv=-\gamma v\,\mathrm dt+\sqrt{2D_v}\,\mathrm dW$ .
Its velocity autocorrelation decays exponentially, $\langle v(t)v(0)\rangle=(k_BT/m)e^{-\gamma|t|}$ , with correlation time $\tau=1/\gamma$ set by the friction.
Its power spectrum is a Lorentzian $S(\omega)=(k_BT/m)\,2\gamma/(\gamma^2+\omega^2)$ — the Wiener-Khinchin transform of the exponential autocorrelation, with corner frequency $\omega_c=\gamma$ .
The stationary distribution is Maxwell-Boltzmann, fixed by the fluctuation-dissipation relation $D_v=\gamma k_BT/m$ between noise strength and friction.
Integrating the velocity recovers Brownian motion: ballistic $\langle x^2\rangle\simeq(k_BT/m)t^2$ at short times, diffusive $\langle x^2\rangle\simeq2Dt$ with $D=k_BT/(m\gamma)$ at long times.
Key idea: a linear restoring drift plus white noise yields an exactly solvable Gaussian Markov process whose exponential autocorrelation $e^{-\gamma|t|}$ and stationary Gaussian width are locked together by the fluctuation-dissipation relation $D_v=\gamma k_BT/m$ ; Doob's theorem makes it the unique such process.

The core insight

A linear restoring drift plus white noise yields an exactly solvable Gaussian Markov process whose exponential autocorrelation $e^{-\gamma|t|}$ and stationary Gaussian width are locked together by the fluctuation-dissipation relation $D_v=\gamma k_BT/m$ ; by Doob's theorem it is the only stationary Gaussian Markov process.

Where this lives today

The Ornstein-Uhlenbeck process is the working model of any physical noise with a finite correlation time. Its defining application — the velocity of a Brownian particle — is now directly observed: high-bandwidth optical-tweezer and particle-tracking experiments resolve the exponential velocity autocorrelation, the equipartition variance $k_BT/m$ , and the long-predicted ballistic regime $\langle x^2\rangle\sim t^2$ at very short times, confirming the crossover this article derives. Because Doob's theorem makes it the unique stationary Gaussian Markov process, it is the default surrogate for colored noise in simulations, where it regularizes the singular delta-correlated white noise into something integrable with a finite time step. The same equation runs far beyond physics. In neuroscience the subthreshold membrane potential of a leaky integrate-and-fire neuron is an OU process; in quantitative finance the Vasicek model of interest rates and many mean-reverting commodity and volatility models are OU processes, prized because their closed-form Gaussian transition density makes likelihoods and option prices analytically tractable. And near equilibrium, any thermodynamic variable that relaxes linearly obeys an OU equation — the decay of its spontaneous fluctuations follows the same law as its macroscopic relaxation, which is Onsager's regression hypothesis and the concrete face of the fluctuation-dissipation theorem and linear response.

The physical picture

Picture a marble in a smooth bowl, resting at the bottom but constantly jostled by a fan of random air puffs. The bowl walls — the linear drift $-\gamma v$ — always push it back toward center; the puffs — white noise — keep knocking it out. The steady-state jitter is the OU stationary distribution, and the time the marble takes to forget a displacement and settle back is the correlation time $\tau=1/\gamma$ . That single picture contains the whole process. The exponential correlation $e^{-\gamma|t|}$ is the fingerprint of a single relaxation rate with no memory beyond the present state — which is exactly what "Markov plus stationary plus Gaussian" demands. Any sharper decay, or any oscillation, would require additional degrees of freedom and would break one of these three properties. This is the content of Doob's theorem, and it is why the OU process is not just a model of colored noise but the canonical one. Two competing tendencies set the variance. Friction alone would relax the velocity to zero (variance $\to0$ ); noise alone would make the variance grow without bound (free diffusion, the Wiener process). Their balance pins the stationary variance at $k_BT/m$ — the equipartition value. Temperature emerges here as the ratio of agitation to dissipation: $D_v/\gamma=k_BT/m$ . This is the fluctuation-dissipation relation in its simplest concrete form, and it is what makes the stationary distribution the Maxwell-Boltzmann Gaussian rather than something arbitrary. Watching the process on different time scales reveals different physics, and this is the entire story of Brownian motion compressed into one autocorrelation function. Faster than $\tau$ the particle moves ballistically and remembers its velocity; slower than $\tau$ the velocity has randomized many times and the displacement performs ordinary diffusion. In the frequency domain the same crossover appears as the Lorentzian corner at $\omega_c=\gamma$ : below it the spectrum is flat (white, at value $2k_BT/m\gamma$ ), above it the spectrum rolls off as $1/\omega^2$ .

Figure

The OU marble in a bowl: linear drift pulls back, white noise kicks out, temperature sets the jitter

Figure

Punchline: exponential autocorrelation ⟷ Lorentzian spectrum, joined at the corner frequency ω_c = γ

The core equations

The process is defined by a linear Langevin stochastic differential equation: a restoring drift proportional to the velocity plus additive white noise. This is the simplest non-trivial stochastic differential equation, and its linearity is what makes everything solvable in closed form.

\mathrm{d}v = -\gamma\,v\,\mathrm{d}t + \sqrt{2D_v}\,\mathrm{d}W(t) \qquad (1)

where: $v(t)$ — velocity of the Brownian particle (the OU variable); $\gamma$ — friction/relaxation rate (units $1/$ time), $\gamma=\zeta/m$ with $\zeta$ the drag coefficient; $D_v$ — velocity-space noise strength; $\mathrm dW$ — Wiener increment, $\langle\mathrm dW\rangle=0$ , $\langle\mathrm dW^2\rangle=\mathrm dt$ . The correlation time is $\tau=1/\gamma$ ; the noise is additive (constant coefficient), so Itô and Stratonovich interpretations coincide. Because the equation is linear with additive noise, it integrates exactly with an integrating factor, giving the velocity as a deterministic decaying transient plus a Gaussian stochastic integral over the past noise.

v(t) = v_0\,e^{-\gamma t} + \sqrt{2D_v}\int_0^t e^{-\gamma(t-s)}\,\mathrm{d}W(s) \qquad (2)

where: $v_0$ — initial velocity; $e^{-\gamma(t-s)}$ — memory kernel weighting past noise. Being a linear functional of Gaussian noise, $v(t)$ is Gaussian for all $t$ ; its mean and variance therefore determine the entire distribution. Taking moments of the exact solution gives the decaying mean and the saturating variance. The variance interpolates from zero (a sharp initial velocity) to the equilibrium value.

\langle v(t)\rangle = v_0 e^{-\gamma t},\qquad \mathrm{Var}[v(t)] = \frac{D_v}{\gamma}\left(1-e^{-2\gamma t}\right) \qquad (3)

where: $D_v/\gamma$ — stationary variance, equal to $k_BT/m$ by fluctuation-dissipation. The mean decays at rate $\gamma$ ; the variance saturates at rate $2\gamma$ — twice as fast, because variance is quadratic in the fluctuation. In the stationary state the two-time velocity autocorrelation is a single decaying exponential — the defining signature of the OU process and the content of Doob's uniqueness theorem.

C(t)=\langle v(t)\,v(0)\rangle_{\rm stat} = \frac{k_BT}{m}\,e^{-\gamma|t|} \qquad (4)

where: $k_BT/m$ — equilibrium velocity variance (one-dimensional equipartition); $|t|$ — absolute lag; $C$ is even and decays at rate $\gamma$ . This exponential is the unique correlation compatible with a stationary Gaussian Markov process (Doob), as shown in the derivation. The noise strength is not free: demanding that the stationary variance equal the equipartition value fixes it in terms of friction and temperature — the fluctuation-dissipation relation, recalled by name from the fluctuation-dissipation-theorem article.

D_v = \gamma\,\frac{k_BT}{m} \qquad (5)

where: $k_BT$ — thermal energy; $m$ — mass. This is the simplest instance of fluctuation-dissipation: the noise amplitude ( $D_v$ ) and the dissipation ( $\gamma$ ) cannot be chosen independently — equilibrium ties them through the temperature. The Fourier transform of the exponential autocorrelation is a Lorentzian power spectrum (Wiener-Khinchin), whose half-width is the corner frequency $\gamma$ .

S(\omega)=\int_{-\infty}^{\infty} C(t)\,e^{-i\omega t}\,\mathrm{d}t = \frac{k_BT}{m}\,\frac{2\gamma}{\gamma^2+\omega^2} \qquad (6)

where: $S(\omega)$ — spectral density of velocity fluctuations; $2\gamma/(\gamma^2+\omega^2)$ — Lorentzian of half-width $\gamma$ . It is flat (white) below $\omega=\gamma$ at value $S(0)=2k_BT/m\gamma$ and rolls off as $1/\omega^2$ above it; the corner marks the crossover between memory and forgetfulness. Integrating the velocity gives Brownian displacement, and the long-time diffusion constant is the Green-Kubo integral of the velocity autocorrelation — the Einstein relation, recalled by name.

D=\int_0^{\infty}\langle v(t)v(0)\rangle\,\mathrm{d}t = \frac{k_BT}{m\gamma} \qquad (7)

where: $D$ — spatial diffusion constant of the integrated process. The time integral of the exponential correlation is $(k_BT/m)/\gamma=k_BT/m\gamma$ : the correlation area is the diffusion constant, the content of the Green-Kubo formula for a single particle.

Derivation AThe velocity autocorrelation and Doob's uniqueness

The goal is to compute the stationary velocity autocorrelation directly from the exact solution and to see why it must be a single exponential — Doob's theorem that the OU process is the unique stationary Gaussian Markov process. The strategy is to write both velocity factors as integrals over the driving noise, average using the Itô isometry (which turns the double noise integral into a single integral over coincident times), evaluate the elementary exponential integral, take the stationary limit, and finally show that the exponential is the only correlation consistent with the Markov (Chapman-Kolmogorov) factorization. Watch the stationary limit: transients die and only the lag-dependent term survives.

Step 1. Write both factors from the exact solution, $v(t)=v_0e^{-\gamma t}+\sqrt{2D_v}\int_0^t e^{-\gamma(t-s)}\mathrm dW(s)$ and similarly $v(t')$ with $t'\ge t$ . Why: the correlation is an average of a product of two Gaussian integrals, and the deterministic transient and the noise integral must be tracked separately.

Step 2. Average. The cross terms between the deterministic part and the zero-mean noise integral vanish, leaving $\langle v(t)v(t')\rangle=v_0^2e^{-\gamma(t+t')}+2D_v\int_0^t\int_0^{t'}e^{-\gamma(t-s)}e^{-\gamma(t'-s')}\langle\mathrm dW(s)\mathrm dW(s')\rangle$ . Why: only the noise autocorrelation survives averaging, and the Itô isometry $\langle\mathrm dW(s)\mathrm dW(s')\rangle=\delta(s-s')\mathrm ds\,\mathrm ds'$ collapses the double integral.

Step 3. Apply the isometry and integrate: the double integral becomes $2D_v e^{-\gamma(t+t')}\int_0^{\min(t,t')}e^{2\gamma s}\mathrm ds=\tfrac{D_v}{\gamma}\big(e^{-\gamma(t'-t)}-e^{-\gamma(t+t')}\big)$ . Why: the delta function forces $s=s'$ , and the remaining single exponential integral is elementary.

Step 4. Take the stationary limit $t,t'\to\infty$ with $|t'-t|$ fixed (or start from the stationary distribution so $\langle v_0^2\rangle=D_v/\gamma$ ): the transients $e^{-\gamma(t+t')}$ die and $\langle v(t)v(t')\rangle\to\tfrac{D_v}{\gamma}e^{-\gamma|t'-t|}=\tfrac{k_BT}{m}e^{-\gamma|t'-t|}$ , using $D_v/\gamma=k_BT/m$ . Why: in the stationary state the correlation depends only on the lag, and the equipartition value fixes its amplitude.

Step 5. Establish Doob's uniqueness. For a stationary Gaussian process the Markov property requires the correlation to satisfy the Chapman-Kolmogorov factorization $C(t_1+t_2)=C(t_1)C(t_2)/C(0)$ ; the only continuous solution of this functional equation is the exponential $C(t)=C(0)e^{-\gamma|t|}$ . Why: "stationary + Gaussian + Markov" is exactly three constraints, and they force a single relaxation rate — the OU process is therefore the unique such process, and any non-exponential correlation (e.g. a Gaussian) fails the factorization and cannot be Markov. *Verified: verify_aq-sm-ornstein-uhlenbeck-process.py confirms $C(0)=k_BT/m$ (equipartition), the evenness of $C$ , and the Chapman-Kolmogorov factorization for the exponential — while a Gaussian correlation fails it, confirming Doob's uniqueness.*

Go deeper · Derivation B · 1 minFrom OU velocity to the Einstein diffusion constant

Derivation BFrom OU velocity to the Einstein diffusion constant

The goal is to integrate the OU velocity into a displacement and show that the mean-square displacement crosses over from ballistic ( $\propto t^2$ ) at short times to diffusive ( $\propto t$ ) at long times, with the long-time diffusion constant $D=k_BT/(m\gamma)$ — the Einstein relation, obtained equivalently as the Green-Kubo integral of the velocity autocorrelation. The strategy is to write the displacement as the time integral of the velocity, express the mean-square displacement as a double integral of the velocity correlation, evaluate it exactly, and expand in the two limits. Watch how the same diffusion constant emerges from both the long-time MSD slope and the correlation-function area.

Step 1. Write $x(t)=\int_0^t v(s)\,\mathrm ds$ , so $\langle x^2(t)\rangle=\int_0^t\int_0^t\langle v(s)v(s')\rangle\,\mathrm ds\,\mathrm ds'=\tfrac{k_BT}{m}\int_0^t\int_0^t e^{-\gamma|s-s'|}\mathrm ds\,\mathrm ds'$ . Why: displacement is the accumulated velocity, and its variance is the double integral of the velocity correlation over the elapsed time.

Step 2. Evaluate the double integral using the symmetry of $|s-s'|$ : $\int_0^t\int_0^t e^{-\gamma|s-s'|}\mathrm ds\,\mathrm ds'=\tfrac{2}{\gamma}\big[t-\tfrac{1-e^{-\gamma t}}{\gamma}\big]$ , giving $\langle x^2(t)\rangle=\tfrac{2k_BT}{m\gamma}\big[t-\tfrac{1-e^{-\gamma t}}{\gamma}\big]$ . Why: splitting the square into $s>s'$ and $s<s'$ and integrating the exponential yields this exact closed form.

Step 3. Expand at short time ( $\gamma t\ll1$ ): $e^{-\gamma t}\approx1-\gamma t+\tfrac12\gamma^2t^2$ , so $\langle x^2(t)\rangle\approx\tfrac{k_BT}{m}t^2$ — ballistic motion at the thermal speed. Why: for times shorter than the correlation time the velocity is nearly constant, so displacement grows linearly in time and its square as $t^2$ .

Step 4. Take the long-time limit ( $\gamma t\gg1$ ): $\langle x^2(t)\rangle\approx\tfrac{2k_BT}{m\gamma}t\equiv2Dt$ , so $D=k_BT/(m\gamma)$ — the Einstein relation. Equivalently, the Green-Kubo integral $D=\int_0^\infty(k_BT/m)e^{-\gamma\tau}\mathrm d\tau=k_BT/(m\gamma)$ gives the identical $D$ . Why: after many correlation times the velocity has decorrelated repeatedly, so the displacement is a sum of independent increments — ordinary diffusion — and the diffusion constant is the area under the velocity correlation. *Verified: verify_aq-sm-ornstein-uhlenbeck-process.py confirms the ballistic short-time limit $\langle x^2\rangle\to(k_BT/m)t^2$ , the diffusive long-time limit $\langle x^2\rangle\to2Dt$ , and that the Green-Kubo integral equals $k_BT/(m\gamma)$ .*

Limiting cases

The OU process interpolates between several familiar idealizations as time scales and parameters vary:

Case	Result	Meaning
Short-time (ballistic)	$\langle[x(t)-x(0)]^2\rangle\simeq\tfrac{k_BT}{m}t^2\ (t\ll\tau)$	For times shorter than the correlation time, velocity is nearly constant and displacement grows ballistically at the thermal speed.
Long-time (diffusive)	$\langle[x(t)-x(0)]^2\rangle\simeq2Dt,\ D=\tfrac{k_BT}{m\gamma}\ (t\gg\tau)$	Velocity decorrelates many times; the integrated OU velocity becomes ordinary Brownian motion with the Einstein diffusion constant.
White-noise / strong-friction	$\langle v(t)v(0)\rangle\to2D\,\delta(t)$ as $\gamma\to\infty$	As $\gamma\to\infty$ with $D=D_v/\gamma^2$ fixed, the exponential correlation collapses to a delta function and $v$ becomes the white-noise force driving overdamped Brownian motion.
Overdamped harmonic well	$\dot x=-(\omega_0^2/\gamma)x+\text{noise}$	For a particle in a harmonic well at large friction, the overdamped position itself becomes an OU process — same mathematics, with relaxation rate $\omega_0^2/\gamma$ .

Worked examples

Example 1 — Equipartition, correlation time, and the zero-frequency spectrum for a Brownian particle. This computes the three numbers that characterize a concrete OU process.

Step 1. Take a Brownian particle with relaxation rate $\gamma=2$ (in units where $k_BT/m=1$ ), so the correlation time is $\tau=1/\gamma=\mathbf{0.5}$ .

Step 2. The stationary velocity variance is $\langle v^2\rangle=k_BT/m=\mathbf{1}$ — equipartition, independent of $\gamma$ . The noise strength required to maintain it is $D_v=\gamma k_BT/m=\mathbf{2}$ (fluctuation-dissipation).

Step 3. The zero-frequency spectral density is $S(0)=2k_BT/(m\gamma)=2\times1/2=\mathbf{1}$ , and the Lorentzian corner (where $S$ falls to half its zero-frequency value) sits exactly at $\omega_c=\gamma=\mathbf{2}$ .

Step 4. So the three characteristic numbers are the correlation time $\tau=0.5$ , the variance $\langle v^2\rangle=1$ , and the spectral corner $\omega_c=2$ — and they are not independent: the corner is the inverse correlation time, and the flat spectral level $S(0)=2\langle v^2\rangle\tau$ ties the spectrum to the variance and correlation time. One relaxation rate $\gamma$ fixes the entire time- and frequency-domain description.

Example 2 — The ballistic-to-diffusive crossover and the diffusion constant. This computes the mean-square displacement at short and long times and extracts $D$ .

Step 1. Use the exact MSD $\langle x^2(t)\rangle=\tfrac{2k_BT}{m\gamma}\big[t-\tfrac{1-e^{-\gamma t}}{\gamma}\big]$ with $k_BT/m=1$ , $\gamma=2$ , so the diffusion constant is $D=k_BT/(m\gamma)=\mathbf{0.5}$ .

Step 2. At short time $t=0.001$ (much less than $\tau=0.5$ ): the MSD is $\approx(k_BT/m)t^2=(1)(0.001)^2=\mathbf{10^{-6}}$ — ballistic, $\propto t^2$ .

Step 3. At long time $t=200$ (much greater than $\tau$ ): the MSD is $\approx2Dt=2(0.5)(200)=\mathbf{200}$ — diffusive, $\propto t$ , with slope $2D=1$ .

Step 4. The crossover happens around $t\sim\tau=0.5$ : below it the particle "remembers" its velocity and moves ballistically; above it the velocity has randomized many times and the motion is ordinary diffusion. The single correlation time $\tau=1/\gamma$ marks the boundary between the two regimes — the whole of Brownian motion in one number.

Origins · 1 minornstein-uhlenbeck process and how this was discovered

ornstein-uhlenbeck process

process introduced by L. S. Ornstein and G. E. Uhlenbeck (1930) · The Ornstein-Uhlenbeck process is named for the Dutch physicists Leonard Salomon Ornstein and George Eugene Uhlenbeck, who introduced it in their 1930 paper "On the theory of the Brownian motion" (Phys. Rev. 36, 823). Their aim was to give the velocity of a Brownian particle a mathematically well-defined description, since the Wiener process (modeling displacement) has non-differentiable paths. By adding linear friction they obtained a smooth, stationary, Gaussian, mean-reverting process with an exponential autocorrelation and a Maxwell-Boltzmann stationary distribution. Its canonical status was cemented by Joseph Doob, who proved in 1942 that it is the unique stationary Gaussian Markov process — making it the fundamental model of colored (finite-correlation-time) noise across physics, engineering, and quantitative finance.

History

1905

— Einstein's theory of Brownian motion relates diffusion to molecular agitation

1908

— Langevin introduces the stochastic force equation for the Brownian particle

1930

— Ornstein and Uhlenbeck give the velocity a well-defined description via linear friction (Phys. Rev. 36, 823)

1942

— Doob proves the OU process is the unique stationary Gaussian Markov process

1945

— Wang and Uhlenbeck's review consolidates the theory of the Brownian motion (Rev. Mod. Phys. 17, 323)

Going further

References. Pathria & Beale — Statistical Mechanics, 3rd ed., Ch. 15 (Fluctuations and stochastic processes) · Kardar — Statistical Physics of Particles, Ch. 9 · Reif — Fundamentals of Statistical and Thermal Physics, Ch. 15 · Sethna — Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 10 · Uhlenbeck & Ornstein — Phys. Rev. 36, 823 (1930) · Wang & Uhlenbeck — Rev. Mod. Phys. 17, 323 (1945) · Risken — The Fokker-Planck Equation, 2nd ed., Ch. 3 · Gardiner — Handbook of Stochastic Methods, Ch. 4.

Provenance. All equations are standard and appear in Wang & Uhlenbeck (Rev. Mod. Phys. 1945), Risken Ch. 3, Gardiner Ch. 4, and Kardar Ch. 9. A verifier recomputes: (1) the integrating-factor solution and the Itô-isometry evaluation of $\langle v(t)v(t')\rangle$ giving $(k_BT/m)e^{-\gamma|t-t'|}$ ; (2) the double-integral MSD $\langle x^2\rangle=(2k_BT/m\gamma)[t-(1-e^{-\gamma t})/\gamma]$ and its ballistic/diffusive limits; (3) the stationary Fokker-Planck solution reducing to Maxwell-Boltzmann under $D_v=\gamma k_BT/m$ ; (4) the Wiener-Khinchin transform of $e^{-\gamma|t|}$ giving the Lorentzian $2\gamma/(\gamma^2+\omega^2)$ with corner $\omega_c=\gamma$ ; (5) the Green-Kubo integral $D=k_BT/(m\gamma)$ ; (6) Doob's Chapman-Kolmogorov factorization for the exponential (and its failure for a Gaussian). Convention: $\gamma$ is the relaxation rate (units $1/$ time), $\tau=1/\gamma$ . No empirical constants are quoted.

Builds on: The Langevin Equation · Wiener Process and White Noise · Brownian Motion · Stochastic Calculus: Itô versus Stratonovich Leads to: The Fokker-Planck Equation · The Fluctuation-Dissipation Theorem · The Einstein Relation · Linear Response Theory · Kramers' Escape Rate Theory

Is the OU process the same as Brownian motion?

Not quite — it is the velocity of a Brownian particle. Standard Brownian motion (the Wiener process) models displacement: its variance grows linearly forever and it has no stationary distribution. The OU process models velocity: it has an exponential autocorrelation and a stationary Gaussian (Maxwell-Boltzmann) distribution. Integrating the OU velocity recovers Brownian displacement, with ballistic motion at short times and ordinary diffusion at long times. So they are intimately related — one is the time integral of the other — but distinct: OU is stationary, Brownian displacement is not.

Why is the autocorrelation exactly exponential and not some other decaying function?

Because of Doob's theorem. A process that is simultaneously stationary, Gaussian, and Markovian must have a correlation obeying the Chapman-Kolmogorov factorization $C(t_1+t_2)=C(t_1)C(t_2)/C(0)$ , and the only continuous solution of this functional equation is the exponential. Any other shape — a Gaussian, a stretched exponential, an oscillation — would violate one of the three properties (usually the Markov property, requiring extra hidden variables). The single exponential is the fingerprint of a single relaxation rate with no memory.

What does the corner frequency of the Lorentzian mean physically?

The corner $\omega_c=\gamma$ separates two regimes. Below it (slow probing, long times) the spectrum is flat and white: the velocity has fully decorrelated, so fluctuations look uncorrelated. Above it (fast probing, short times) the spectrum rolls off as $1/\omega^2$ : the velocity still "remembers" itself on these timescales, so high-frequency fluctuations are suppressed. The corner is the inverse of the correlation time, $\omega_c=1/\tau$ , and marks the crossover from memory to forgetfulness — the same crossover that appears in the time domain as ballistic-to-diffusive motion.

How does the fluctuation-dissipation relation appear in the OU process?

It appears as the constraint $D_v=\gamma k_BT/m$ tying the noise strength $D_v$ to the friction $\gamma$ . If you chose them independently, the stationary variance $D_v/\gamma$ would be arbitrary and the equilibrium distribution would not be Maxwell-Boltzmann. Demanding equipartition, $\tfrac12 m\langle v^2\rangle=\tfrac12 k_BT$ , forces $D_v/\gamma=k_BT/m$ . This is the simplest concrete instance of the general fluctuation-dissipation theorem: the same molecular collisions that dissipate energy (friction) also drive the fluctuations (noise), and equilibrium fixes their ratio through the temperature.

Can the OU process describe a particle in a harmonic well, not just free velocity?

Yes — in the overdamped limit. For a particle in a harmonic potential $U=\tfrac12 m\omega_0^2 x^2$ at high friction, the position obeys $\dot x=-(\omega_0^2/\gamma)x+\text{noise}$ , which is exactly an OU equation with the position as the OU variable and relaxation rate $\omega_0^2/\gamma$ . The stationary distribution is then the Boltzmann distribution in the well, $P(x)\propto e^{-m\omega_0^2 x^2/2k_BT}$ . So the OU process describes both the free Brownian velocity and the overdamped position in a harmonic trap — the same mathematics, different physical variable.

Why is the OU process so useful as a numerical model of noise?

Two reasons. First, real physical noise always has a finite correlation time, and the OU process is the minimal (unique, by Doob) stationary Gaussian model with one. Second, it regularizes white noise: the idealized delta-correlated white noise is singular and awkward to integrate with a finite time step, but the OU process replaces the delta with an exponential of finite width $\tau$ , giving smooth, integrable trajectories that converge to white noise as $\tau\to0$ . Its closed-form Gaussian transition density also makes exact updates possible in a simulation, avoiding discretization error in the noise itself.