Drift Nuisance And Diffusion Estimation

Why Drift Nuisance Can Matter Less at High Frequency

A diffusion model has two distinct sources of short-term change: a systematic drift and random variation whose scale is controlled by the diffusion coefficient. When observations become closer together, a bounded drift changes the process by an amount proportional to the time interval, whereas Brownian variation is proportional to the square root of that interval. This difference in scale is the basis for estimating diffusion while treating drift parameters as nuisance parameters.

The conclusion is conditional, not universal. It relies on a fixed time horizon, increasingly fine observations, a correctly specified continuous-path diffusion, and drift that does not grow with observation frequency. It can fail, for example, when measurements contain independent observation error or when a supposed drift term becomes as large as the Brownian increment at fine scales.

Definition: Consider a scalar diffusion observed at times $t_i=i\Delta$, $i=0,\ldots,n$, over a fixed horizon $T=n\Delta$. In the basic setting, $$ dX_t=b(X_t,\beta)\,dt+\sigma\,dW_t,\qquad 0\leq t\leq T, $$ where $W_t$ is Brownian motion, $\beta$ is a drift parameter, and $\sigma>0$ is a constant diffusion coefficient. Write $v=\sigma^2$ for the diffusion variance. High-frequency sampling means $\Delta\to0$ and $n\to\infty$ while $T$ stays fixed. The parameter $\beta$ is a nuisance parameter for diffusion estimation when the target is $v$, not $\beta$.

For the first result, assume the process exists on $[0,T]$, and that the drift along its path is bounded: $|b(X_t,\beta)|\leq B$ almost surely for a finite constant $B$. The bound is uniform over the parameter values under consideration. These assumptions give a direct way to quantify the contribution of drift to squared increments.

Quadratic Variation Separates the Scales

Define the realized quadratic-variation estimate of $v$ by $Q_n=T^{-1}\sum_{i=1}^n(X_{t_i}-X_{t_{i-1}})^2$. For a continuous diffusion with constant $\sigma$, quadratic variation over $[0,T]$ is $vT$. The estimate $Q_n$ approximates that quantity after division by $T$. The following argument shows explicitly why bounded drift does not change its probability limit.

Theorem: Under the bounded-drift assumptions above, with continuous observation of $X$ at the grid points and no measurement error, $Q_n\xrightarrow{p}v$ as $\Delta\to0$ at fixed $T$.

Split each increment.
Write $\Delta_iX=X_{t_i}-X_{t_{i-1}}=A_i+\sigma Z_i$, where $A_i=\int_{t_{i-1}}^{t_i}b(X_s,\beta)\,ds$ and $Z_i=W_{t_i}-W_{t_{i-1}}$. The drift bound gives $|A_i|\leq B\Delta$.

Separate the squared terms.
Expanding $Q_n$ gives $Q_n=\frac{\sigma^2}{T}\sum_i Z_i^2+\frac{2\sigma}{T}\sum_i Z_iA_i+\frac{1}{T}\sum_i A_i^2$.

Control drift squared.
$0\leq T^{-1}\sum_i A_i^2\leq T^{-1}nB^2\Delta^2=B^2\Delta$, which tends to zero.

Control the cross term.
Using the triangle inequality and $E|Z_i|=\sqrt{2\Delta/\pi}$, $E|2\sigma T^{-1}\sum_i Z_iA_i|\leq 2\sigma BT^{-1}n\Delta\sqrt{2\Delta/\pi}=2\sigma B\sqrt{2\Delta/\pi}$. This tends to zero, so the cross term tends to zero in probability by Markov's inequality.

Take the Brownian limit.
The $Z_i$ are independent $N(0,\Delta)$, so $E[T^{-1}\sum_iZ_i^2]=1$ and $\operatorname{Var}(T^{-1}\sum_iZ_i^2)=2/n$. Thus this term converges in probability to $\sigma^2=v$. Combining the three terms proves $Q_n\xrightarrow{p}v$. $\square$

The argument establishes consistency, not that the drift is irrelevant in every finite sample. Nor does it establish that drift parameters themselves become easy to estimate: at fixed $T$, finer sampling does not create a longer time series. It establishes that bounded drift contributes a vanishing amount to this particular high-frequency diffusion estimate.

Worked Example: Quantifying a Bounded Drift Contribution

Suppose $dX_t=0.8\,dt+0.5\,dW_t$ over $T=1$, so the true diffusion variance is $v=0.25$. Under equal intervals, the expected uncentered estimate $Q_n$ can be calculated exactly.

For each interval, $\Delta_iX=0.8\Delta+0.5Z_i$, with $E[Z_i]=0$ and $E[Z_i^2]=\Delta$. Therefore $E[(\Delta_iX)^2]=(0.8)^2\Delta^2+(0.5)^2\Delta=0.64\Delta^2+0.25\Delta$. Summing $n$ intervals and dividing by $T=1$ gives $E[Q_n]=0.25+0.64n\Delta^2=0.25+0.64/n$, since $\Delta=1/n$.

At $n=100$, $E[Q_n]=0.25+0.0064=0.2564$. At $n=400$, $E[Q_n]=0.25+0.0016=0.2516$. The drift contribution to this expectation shrinks as $1/n$. This calculation is not a claim that an individual estimate must be close to its expectation; Brownian sampling variability remains.

Likelihood Shows Different Information Rates

The same separation is visible in an exact Gaussian increment model. Suppose $b(X_t,\beta)=\beta$, so the drift is constant. Then the increments are independent and $\Delta_iX\sim N(\beta\Delta,v\Delta)$. The likelihood uses both parameters jointly, but its information about them accumulates at different rates.

Result: In the constant-drift Gaussian model with known initial value and fixed $T=n\Delta$, the Fisher information matrix for $(\beta,v)$ is diagonal, with entries $T/v$ and $n/(2v^2)$. Thus drift information remains fixed as $\Delta\to0$, while diffusion information grows in proportion to $n$.

To derive this, let $e_i=\Delta_iX-\beta\Delta$, so $e_i\sim N(0,v\Delta)$. Apart from terms not depending on the parameters, the log-likelihood is $\ell(\beta,v)=-\frac n2\log(v\Delta)-\frac{1}{2v\Delta}\sum_i e_i^2$. The score for drift is $S_\beta=\sum_i e_i/v$, and the score for diffusion variance is $S_v=\sum_i(e_i^2-v\Delta)/(2v^2\Delta)$. Since a centered normal variable has zero third moment, $E[S_\beta S_v]=0$. Also, $E[e_i^2]=v\Delta$ and $\operatorname{Var}(e_i^2)=2v^2\Delta^2$. Independence of increments therefore gives $E[S_\beta^2]=n\Delta/v=T/v$, $E[S_v^2]=n/(2v^2)$, and zero cross-information. This proves the stated information matrix. $\square$

This exact orthogonality is a property of the constant-drift Gaussian increment model; it should not be assumed for every nonlinear diffusion or every approximate likelihood. The broader scale distinction is useful, but the precise likelihood calculations depend on the model and observation scheme.

Worked Example: Joint Gaussian Estimates From Four Increments

Take $T=1$, $n=4$, $\Delta=0.25$, and observed increments $0.30,-0.10,0.25,0.05$. Under the constant-drift model, the maximum-likelihood drift estimate is the total change divided by elapsed time: $\hat\beta=(0.30-0.10+0.25+0.05)/1=0.50$.

The fitted drift increment on each interval is $\hat\beta\Delta=0.50(0.25)=0.125$. The residual increments are $0.175,-0.225,0.125,-0.075$, whose squared values sum to $0.030625+0.050625+0.015625+0.005625=0.1025$.

The Gaussian maximum-likelihood estimate of the diffusion variance is residual sum of squares divided by $n\Delta$: $\hat v=0.1025/(4(0.25))=0.1025$. For comparison, the uncentered quadratic-variation estimate is $Q_n=(0.30^2+(-0.10)^2+0.25^2+0.05^2)/1=0.165$. Here drift adjustment changes the estimate appreciably. With only four intervals, the high-frequency asymptotic argument does not promise a negligible finite-sample difference.

Where the Separation Fails

The assumptions matter in practice. A key diagnostic is whether observed increments behave like a drift contribution of order $\Delta$ plus diffusion variation of order $\sqrt{\Delta}$. If the measurement process or model violates that scale structure, refining the sampling grid may not improve diffusion estimation.

Worked Example: Independent Measurement Error Dominates Fine Increments

Suppose the latent process satisfies $dX_t=\beta\,dt+0.5\,dW_t$, but observations are $Y_{t_i}=X_{t_i}+\epsilon_i$, where the independent errors have mean zero and variance $\tau^2=0.01$. A measured increment is $\Delta_iY=\Delta_iX+\epsilon_i-\epsilon_{i-1}$.

Independence gives $E[(\epsilon_i-\epsilon_{i-1})^2]=2\tau^2=0.02$, and the cross term with the latent increment has expectation zero. Hence, for the uncentered realized-variation estimate based on $Y$, $E[T^{-1}\sum_i(\Delta_iY)^2]=v+\beta^2T/n+2n\tau^2/T$.

With $T=1$, $\beta=0.8$, and $n=100$, this expectation is $0.25+0.0064+2=2.2564$, rather than a value near $0.25$. As $n$ increases, the measurement-error contribution $2n\tau^2/T$ increases. The latent diffusion’s high-frequency signal has not become easier to read: independent errors are being differenced repeatedly. A method that explicitly accounts for observation error, or data collected under a suitable observation design, is needed instead of applying the noise-free quadratic-variation argument.

A second failure occurs if the drift is not uniformly bounded as sampling gets finer. For example, if a sequence of models has drift $\beta_\Delta=\kappa/\sqrt{\Delta}$, then its mean increment is $\beta_\Delta\Delta=\kappa\sqrt{\Delta}$, the same order as a Brownian increment. In the constant-drift model, the expected squared increment divided by $\Delta$ is $v+\kappa^2$, not $v$. The usual fixed-parameter, bounded-drift result does not apply.

Jumps also invalidate the continuous-path quadratic-variation target: squared increments can include squared jump sizes, not only the continuous diffusion variance. In applications, examine estimates across sampling intervals and look for sensitivity to the finest grid, unusually large increments, or residual structure inconsistent with the assumed observation model. Such checks do not prove the assumptions, but instability is a warning against interpreting a realized-variation estimate as diffusion alone.

Choosing an Estimation Approach

Quadratic variation is attractive when the process is observed without error at high frequency and the target is a constant diffusion variance: it does not require estimating the drift first. A joint transition likelihood instead uses the assumed increment distribution to estimate drift and diffusion together; it can be more efficient in a correctly specified model and is natural with finite, irregularly spaced observations. Its results can be sensitive to transition-density approximation and model misspecification. Estimating drift, subtracting its fitted increment, and then computing squared residuals is another practical approach, but it does not guarantee that drift uncertainty is negligible at a given sample size.

These approaches answer different questions. High-frequency separation is a statement about asymptotic scales under stated assumptions, not a universal license to omit drift from a finite-sample analysis. Neither quadratic variation nor a Gaussian transition likelihood automatically resolves measurement error, jumps, or an incorrect diffusion model.

Common pitfall: “The drift increment is smaller than the diffusion increment at high frequency, so drift can always be ignored.” This confuses a conditional asymptotic result with a general rule. The correction is to check fixed-horizon sampling, bounded drift, continuous paths, and measurement quality. Even when the result applies, finite-sample drift adjustment can matter, and other sources of variation can dominate.

Check Your Understanding

Work through the questions, then compare your reasoning with the solutions below.

For fixed $T$, how do bounded drift increments and Brownian increments scale with $\Delta$?
In the constant-drift Gaussian model, does drift information grow with $n$ at fixed $T$? What about diffusion-variance information?
With $T=2$, $n=20$, and bounded drift $|b|\leq 3$, what upper bound does the proof give for $T^{-1}\sum_i A_i^2$?
Why can independent observation error make the finest-grid quadratic-variation estimate worse rather than better?

Solutions. (1) The drift increment is at most $B\Delta$, while Brownian variation is of order $\sqrt{\Delta}$; their ratio tends to zero for bounded $B$. (2) Drift information is $T/v$, fixed as $n$ increases; diffusion-variance information is $n/(2v^2)$, which increases with $n$. (3) Here $\Delta=T/n=0.1$, so the bound is $B^2\Delta=9(0.1)=0.9$. (4) Differencing independent observation errors contributes variance $2\tau^2$ per increment. After summing and dividing by $T$, its contribution is $2n\tau^2/T$, which grows with $n$ at fixed $T$.