Why Short Trajectories Need a Correction
A diffusion parameter describes the scale of random variation over time. With only a few observed increments, however, an estimate of that scale can be noticeably affected by how the drift is handled. Centering increments around an estimated drift uses information from the same small data set; dividing by the number of increments rather than accounting for that estimation typically makes the estimated variance too small. A second issue is that an increment observed over a long interval need not have variance equal to the instantaneous diffusion variance multiplied by the interval length.
This tutorial develops exact finite-sample corrections in two settings: constant drift with constant diffusion, and a linear Ornstein–Uhlenbeck (OU) diffusion with known mean-reversion parameters. The results depend on explicit assumptions. They are not universal corrections for every pharmacokinetic SDE: for a nonlinear drift or state-dependent diffusion observed sparsely, the Euler increment approximation may be inaccurate, and there is no automatic degrees-of-freedom adjustment that repairs that model error.
Constant Drift: Estimating the Diffusion from Increments
For equal intervals \(\Delta t_i=\Delta\), the increments are independent normal observations with mean \(b\Delta\) and variance \(\sigma^2\Delta\). Estimating that variance requires removing the unknown common mean. The ordinary sample variance uses denominator \(n-1\), not \(n\). Dividing by \(\Delta\) then estimates the continuous-time diffusion variance \(\sigma^2\).
Proof. Set \(Y_i=\Delta X_i\), so \(Y_i\) are independent \(N(b\Delta,\sigma^2\Delta)\). Let \(Y=(Y_1,\ldots,Y_n)^\mathsf{T}\), and let \(P=I-\mathbf{1}\mathbf{1}^\mathsf{T}/n\), the matrix that subtracts the sample mean. It is symmetric and idempotent, has rank \(n-1\), and annihilates the constant vector \(\mathbf{1}\). Thus \[ \sum_{i=1}^n(Y_i-\overline{Y})^2=Y^\mathsf{T}PY. \] Writing \(Y=b\Delta\mathbf{1}+\sigma\sqrt{\Delta}Z\), with \(Z\sim N(0,I)\), gives \(Y^\mathsf{T}PY=\sigma^2\Delta Z^\mathsf{T}PZ\), since \(P\mathbf{1}=0\). An orthogonal change of coordinates diagonalizes \(P\) into \(n-1\) ones and one zero. The transformed standard normal coordinates remain independent standard normals, so \(Z^\mathsf{T}PZ\) is a sum of squares of \(n-1\) independent standard normal variables: it has a \(\chi^2_{n-1}\) distribution and expectation \(n-1\). Substitution proves both claims. \(\square\)
The denominator correction is also visible in the maximum-likelihood estimate. With the mean estimated, the Gaussian likelihood uses the residual sum of squares divided by \(n\Delta\); its expectation is \((n-1)\sigma^2/n\). The likelihood estimate is therefore biased downwards in finite samples, even though it is consistent as \(n\) increases.
Worked Example: Correcting a Short Equal-Interval Trajectory
Suppose four increments are \(0.2,-0.1,0.3,0.0\) concentration units, each over \(\Delta=0.5\) time units. Assume the constant-drift model and no measurement error. Calculate the corrected estimate of \(\sigma^2\), compare it with the uncorrected likelihood estimate, and form a 95% confidence interval.
The mean increment is \(\overline{\Delta X}=(0.2-0.1+0.3+0.0)/4=0.1\). The residuals are \(0.1,-0.2,0.2,-0.1\), and their squared sum is \(0.01+0.04+0.04+0.01=0.10\). Therefore, \[ \widehat{\sigma}^2=\frac{0.10}{(4-1)(0.5)}=\frac{0.10}{1.5}=0.0667. \] The likelihood estimate divides by \(4(0.5)=2\), giving \(0.10/2=0.0500\). Its value is \(3/4\) of the unbiased estimate, as expected when one mean has been estimated from four increments.
For a normal sample, the theorem gives \((3)(0.5)\widehat{\sigma}^2/\sigma^2\sim\chi^2_3\). Using the 2.5% and 97.5% quantiles \(0.2158\) and \(9.348\) of \(\chi^2_3\), respectively, a 95% interval for \(\sigma^2\) is \[ \left(\frac{0.10}{9.348},\frac{0.10}{0.2158}\right)=(0.0107,0.4634). \] This interval is wide: four increments provide little information about the diffusion scale. Unbiasedness does not imply precision.
Unequal Observation Intervals
When interval lengths differ, raw increments have different variances. Under the same constant-drift model, divide each increment by \(\sqrt{\Delta t_i}\). The resulting response has variance \(\sigma^2\), while its mean is \(b\sqrt{\Delta t_i}\). Estimating the one drift coefficient is a regression through the origin with known predictors \(\sqrt{\Delta t_i}\).
Proof. In vector notation \(y=bx+\sigma Z\), where \(Z\sim N(0,I)\). The fitted residual vector is \(My\), where \(M=I-xx^\mathsf{T}/(x^\mathsf{T}x)\). This matrix is symmetric and idempotent, has rank \(n-1\), and satisfies \(Mx=0\). Hence \(My=\sigma MZ\), and the residual sum of squares is \(\sigma^2 Z^\mathsf{T}MZ\). An orthogonal transformation diagonalizing \(M\) shows that this is \(\sigma^2\) times a sum of \(n-1\) independent squared standard normal variables. Its expectation is \((n-1)\sigma^2\), proving the result. \(\square\)
Worked Example: Unequal Intervals and Weighted Residuals
Take three interval lengths \(1,4,1\), with increments \(0.5,1.2,-0.1\), respectively. Under the model above, calculate the unbiased estimate.
The transformed responses are \(y=(0.5,1.2/2,-0.1)=(0.5,0.6,-0.1)\), and the predictors are \(x=(1,2,1)\). The drift estimate is \[ \widehat b=\frac{1(0.5)+2(0.6)+1(-0.1)}{1^2+2^2+1^2}=\frac{1.6}{6}=0.2667. \] The fitted responses are \((0.2667,0.5333,0.2667)\), so the residuals are \((0.2333,0.0667,-0.3667)\). Their squared sum is \(0.05444+0.00444+0.13444=0.19333\). With two residual degrees of freedom, \[ \widehat{\sigma}^2=0.19333/2=0.09667. \] Using the unweighted variance of the raw increments would ignore their different interval lengths and would not have the stated finite-sample distribution.
Correcting an Exact OU Transition Variance
For sparse observations, using increment variance \(\sigma^2\Delta\) can be a poor approximation when the drift pulls the process back toward a mean. Consider \(dX_t=-\kappa(X_t-m)\,dt+\sigma\,dW_t\), with \(\kappa>0\). Its exact transition over interval \(\Delta\) is \[ X_{t+\Delta}=m+e^{-\kappa\Delta}(X_t-m)+\varepsilon,\qquad \varepsilon\sim N\left(0,\frac{\sigma^2(1-e^{-2\kappa\Delta})}{2\kappa}\right). \] This variance follows by solving the linear SDE: the random part is \(\sigma\int_t^{t+\Delta}e^{-\kappa(t+\Delta-s)}\,dW_s\); its variance is \(\sigma^2\int_0^\Delta e^{-2\kappa u}\,du=\sigma^2(1-e^{-2\kappa\Delta})/(2\kappa)\).
Worked Example: From OU Residual Variance to Diffusion Scale
Suppose \(\kappa=0.2\), \(\Delta=2\), and \(m,\kappa\) are known. For four transitions, the exact conditional residuals are \(0.1,-0.2,0.3,-0.1\). Calculate the diffusion variance estimate.
The residual squared sum is \(0.01+0.04+0.09+0.01=0.15\). Since the transition mean is specified, no mean is estimated from these residuals; the unbiased estimate of their variance \(q\) is \(0.15/4=0.0375\). The exact transition formula gives \[ \widehat{\sigma}^2=\frac{2\kappa\widehat q}{1-e^{-2\kappa\Delta}} =\frac{0.4(0.0375)}{1-e^{-0.8}} =\frac{0.015}{0.5507}\approx0.02724. \] Treating the transition variance as \(\sigma^2\Delta\) would instead give \(0.0375/2=0.01875\). The difference is material at this interval length. The exact chi-square result applies here only because the conditional residuals are independent normal variables with a common variance and the transition mean parameters are known.
Diagnostics, Alternatives, and Limits
Before applying a correction, check that its assumptions match the observations and model. The exact constant-drift results require a continuous, error-free observation of the state, a constant drift over each interval, constant diffusion, and independent Brownian increments. The OU result requires the stated exact transition and known conditional mean parameters for the simple residual-variance calculation. If parameters in the conditional mean are estimated, residual degrees of freedom and the exact distribution require a separate analysis; blindly substituting \(n-1\) is not justified.
| Approach | What it estimates | Key caution |
|---|---|---|
| Corrected increment variance | Diffusion variance under constant drift, with the drift estimated from increments | Exact only under the Gaussian constant-drift assumptions |
| Likelihood variance with divisor \(n\) | Maximum-likelihood variance in the equal-interval normal model | Finite-sample downward bias when the mean is estimated |
| Exact OU transition variance | Continuous-time diffusion scale through the exact transition variance | Requires the OU transition model; estimating its mean parameters changes the uncertainty calculation |
| Euler increment approximation | Local approximation to a general SDE increment | Can be inaccurate for long intervals, strong drift, or rapidly changing diffusion |
A further failure mode is observation error. If observed concentrations equal latent concentrations plus independent measurement errors, differencing observations adds measurement-error variation to increments; that variation can be mistaken for process diffusion. Likewise, state-dependent diffusion makes the increment variance depend on the path state, so a single pooled sample variance need not estimate the intended diffusion scale. These problems call for a model that represents the observation process or state dependence, rather than a mechanical degrees-of-freedom correction.
For a short trajectory, report more than the point estimate: state the increment model and interval lengths, specify whether drift parameters were estimated, and give an uncertainty interval or distribution when its assumptions hold. A very wide interval is a useful warning that the data weakly determine the diffusion scale. Residual patterns or obvious dependence between successive increments also warn against treating the increments as independent Gaussian observations. Finite-sample correction removes a specific bias; it cannot repair misspecification, sparse-data approximation error, or lack of information.
Check Your Understanding
Work each problem, then compare with the solutions below.
- For six equal-interval increments under unknown constant drift, what denominator gives an unbiased estimate of the increment variance after estimating the mean?
- If the corrected estimate is \(\widehat{\sigma}^2\) and the interval length is \(\Delta\), what is the chi-square pivot for the constant-drift model?
- Why should raw increments not be pooled with an ordinary sample variance when observation intervals differ?
- For an OU transition with known parameters, state the exact transition variance and the factor that converts it to \(\sigma^2\).
- Name one reason the constant-drift finite-sample correction may fail for observed pharmacokinetic concentrations.
Solutions. (1) Use \(n-1=5\) for the increment sample variance, then divide by \(\Delta\) to estimate \(\sigma^2\). (2) The pivot is \((n-1)\Delta\widehat{\sigma}^2/\sigma^2\sim\chi^2_{n-1}\). (3) Their variances are \(\sigma^2\Delta t_i\), so the increments are not identically distributed; scale by \(\sqrt{\Delta t_i}\) and fit the corresponding drift predictor. (4) The transition variance is \(q=\sigma^2(1-e^{-2\kappa\Delta})/(2\kappa)\), hence \(\sigma^2=2\kappa q/(1-e^{-2\kappa\Delta})\). (5) Examples include measurement error contaminating increments, state-dependent diffusion, or a nonlinear drift over sparse intervals.