Drift Estimation for Stochastic Differential Equations with Denoising Diffusion Models Marcos Tapia Costa, Nikolas Kantas, George Deligiannidis Abstract We study the estimation of time-homogeneous drift functions in multivariate stochastic differential equations with known diffusion coefficient, from multiple trajectories observed at high frequency over a fixed time horizon. We formulate drift estimation as a denoising problem conditional on previous observations, and propose an estimator of the drift function which is a by-product of training a conditional diffusion model capable of simulating new trajectories dynamically. Across different drift classes, the proposed estimator was found to match classical methods in low dimensions and remained consistently competitive in higher dimensions, with gains that cannot be attributed to architectural design choices alone. 1 1 footnotetext: Department of Mathematics, Imperial College London, UK 2 2 footnotetext: Department of Statistics, University of Oxford, UK 1 Problem Statement Suppose one is interested in modelling a time series as a continuous-time stochastic differential equation (SDE), dY_{t}=\mu(Y_{t})dt+\sigma dB_{t},t\in[0,T], (1) where (B_{t})_{t\in[0,T]} is a D -dimensional Brownian motion and \mu:\mathbbm{R}^{D}\rightarrow\mathbbm{R}^{D} is a time-homogeneous drift function. We assume \sigma is known or well-estimated, i.e., using the quadratic variation of a single path, see ( Jacod 1994 ; Barndorff-Nielsen & Shephard 2002 ; Jacod & Protter 2012 ) . We also assume \mu is locally Lipschitz to ensure a strong solution to ( 1 ) up to terminal time T , see ( Karatzas & Shreve 2000 ) . We will investigate the case we are given a dataset \mathcal{D} consisting of I\gg 1 i.i.d. trajectories, each with J observations of dimension D observed at times t_{j}=j\Delta\ \forall\ j\in\{1,\cdots,J\} , such that \mathcal{D}=\{Y_{t_{1}}^{(i)},\cdots,Y_{t_{J}}\}_{i\in[I]},[I]:=\{1,\cdots,I\} and \Delta:=t_{j+1}-t_{j} is small. In our setting, the number of trajectories I and the number of discrete observations J are fixed with a constant sampling frequency \Delta . Without loss of generality, we assume T=1 and Y_{0}^{(i)}\sim Q_{0} , where Q_{0} admits a density with respect to the Lebesgue measure dy_{0} ; Q_{0}(dy_{0})=q_{0}(y_{0})dy_{0} . We will study the problem of estimating the drift function \mu from the dataset \mathcal{D} . 1.1 Drift estimation as a regression Letting Z_{t_{j}}:=Y_{t_{j}}-Y_{t_{j-1}} , the solution over s\in[t_{j-1},t_{j}] to ( 1 ) takes the form, Z_{t_{j}}=\int_{t_{j-1}}^{t_{j}}\mu(Y_{s})ds+\sqrt{\Delta}\ \omega,\ \omega\sim\mathcal{N}(0,I). (2) For small \Delta , an Euler-Maruyama (EM) approximation to ( 2 ) above yields, \widehat{Z}_{t_{j}}=\mu(Y_{t_{j-1}})\Delta+\sqrt{\Delta}\omega,\omega\sim\mathcal{N}(0,I) (3) and the drift enters \mu through the conditional mean: \mathbbm{E}\left[\widehat{Z}_{t_{j}}\mid Y_{t_{j-1}}\right]=\mu(Y_{t_{j-1}})\Delta. (4) In our multiple trajectory setting, with I\gg 1 , one could use ( 4 ) to learn \mu by regressing states against increments through a nonparametric function D_{\theta} ( Zhao et al. 2025 ) , D_{\theta^{*}}=\arg\min_{\theta}\frac{1}{2}\sum_{i,j}\left[\left\lVert D_{\theta}(Y^{(i)}_{t_{j-1}})-Z^{(i)}_{t_{j}}\right\rVert_{2}^{2}\right]. (5) It is well known that in our fixed T,J,\Delta setting, the drift \mu cannot be estimated consistently from a single trajectory ( Jakobsen & Sørensen 2015 ) . However, in the multiple trajectory setting, ( Zhao et al. 2025 ) provide asymptotic convergence guarantees for separable drift functions under conditions on the class of functions D_{\theta} . Yet, the drift estimator D_{\theta^{*}} can suffer from high variance, as the drift contribution scales as \mathcal{O}(\Delta) , and the diffusion noise scales as \mathcal{O}(\sqrt{\Delta}) ( Gobet 2002 ; Kutoyants 2004 ) , and the Fisher information per observation vanishes as \Delta\rightarrow 0 . In high dimensions, this difficulty is further compounded by deteriorating nonparametric regression convergence rates with dimension, requiring significantly more data to achieve the same accuracy as in lower dimensions ( Tsybakov 2009 ) . A general approach to mitigate variance in regression problems is to inject additive Gaussian noise into the inputs, which acts as a form of Tikhonov regularisation on D_{\theta} ( Bishop 1995 ) . More generally, under the loss in ( 5 ), regression with noisy inputs corresponds to learning the conditional expectation of the target given the corrupted input, i.e., learning a denoiser . Moreover, ( Starnes & Webster 2025 ) show learning the denoiser from Gaussian perturbations can smooth low variance, high curvature directions during optimisation and lead to a better conditioned stochastic gradient descent problem. ( Vincent 2011 ) further showed that such denoising objectives are equivalent to score matching under Gaussian perturbations. Modern denoising diffusion models extend this principle by introducing a controlled, multi-scale noising process and learning denoisers across noise levels, with great success in image and video applications ( Sohl-Dickstein et al. 2015 ; Ho et al. 2020 ; Song et al. 2021b ) . In this paper, we extend the methodology of denoising diffusion models to introduce a controlled noising process and derive an estimator for the drift \mu . We will propose a novel construction for the estimator, and present systematic empirical results showing that the denoising-based estimator improves out-of-sample performance in high-dimensional settings with strong interactions or chaotic dynamics, while remaining competitive when standard regression already generalises well. Theoretical guarantees as in ( Zhao et al. 2025 ) are left for future work. 2 Related Literature 2.1 Drift estimation for stochastic differential equations The problem of estimating the drift function of an SDE has been extensively studied. In the single trajectory setting, ( Kessler 1997 ) derived an efficient estimator for the drift function of an ergodic scalar diffusion observed at discrete times when transition densities are unknown. ( Shoji & Ozaki 1995 ) developed a pseudo-likelihood estimator from a locally linear approximation of the SDE, while ( Aït-Sahalia 2002 ) used polynomial Hermite expansions of the transition density to obtain an approximate maximum likelihood estimator. Nonparametric approaches were introduced through the kernel regression estimator of ( Watson 1964 ) , and ( Nadaraya 1964 ) was first applied to Brownian SDEs by ( Florens-Zmirou 1993 ) , who estimate both drift and diffusion coefficients in scalar diffusions, and ( Bandi & Phillips 2003 ; Bandi & Phillips 2007 ) provided uniform convergence rates for its application to high frequency financial data. Furthermore, ( Dalalyan 2006 ) derived a nonparametric estimator for homogeneous drift functions that achieves minimax optimal rates, while ( Comte & Genon-Catalot 2009 ; Comte & Genon-Catalot 2010 ) proposed penalised least squares projection methods to jointly estimate the drift and diffusion coefficients nonparametrically. More recently, ( Oga & Koike 2023 ) develop theoretical guarantees for multidimensional drift estimation for a class of deep neural networks, given a single path with many observations. In the multiple trajectory setting, which is the observation regime of this paper, ( Bonalli & Rudi 2025 ) used a Reproducing Kernel Hilbert Space approximation of the Fokker-Plank equation to estimate multidimensional non-linear drift functions. ( Mohammadi et al. 2024 ) use Functional Data Analysis to develop a drift estimator given sparse observations. ( Marie & Rosier 2023 ) extend kernel smoothing estimators of the drift function to the multiple i.i.d. trajectory setting, while ( Comte & Genon-Catalot 2020 ; Denis et al. 2021 ) extend penalised least squares estimators to the same setting. More recently, ( Zhao et al. 2025 ) provide asymptotic convergence guarantees for separable drift estimation in a compact domain using feed-forward neural networks. Their empirical evaluation, however, is restricted to a single test drift function and assesses recovery only along one coordinate of a high-dimensional system, leaving performance on fully coupled dynamics unexamined. 2.2 Score-based generative models Score-based generative models, or Denoising Diffusion Models (DDMs) , were introduced in ( Ho et al. 2020 ) and ( Song et al. 2021b ) , drawing ideas from score-matching ( Hyvarinen & Dayan 2005 ; Vincent 2011 ) and building on earlier work ( Sohl-Dickstein et al. 2015 ) . By using neural networks to estimate the log density of a smoothed data distribution, i.e. the score , they have achieved state-of-the-art results in image generation ( Ho et al. 2020 ; Dhariwal & Nichol 2021 ; Saharia et al. 2022 ; Brooks et al. 2023 ) , and have found numerous applications in video generation ( Ho et al. 2022 ) , molecule and drug design, ( Xu et al. 2023 ; Corso et al. 2023 ) , strategy testing ( Koshiyama et al. 2021 ) , and data sharing ( Assefa 2020 ) . Their state-of-the-art performance across several domains, and the fact that they are less susceptible to training instabilities or mode-collapse, have established DDMs as the gold-standard in generative modelling ( Dhariwal & Nichol 2021 ) . In particular, ( Song & Ermon 2019 ) extended ( Sohl-Dickstein et al. 2015 ) to a practical setting by formalising Stochastic Gradient Langevin Dynamics (SGLD) as the latent variable process, while ( Ho et al. 2020 ) introduced Denoising Diffusion Probabilistic Models (DDPM). These approaches were unified under an SDE-based framework in ( Song et al. 2021c ) , who show that the forward processes in DDPM and SGLD are discretisations of continuous-time Itô-SDEs. Further generalisations have been achieved in recent years, see ( Song et al. 2021a ; Franzese et al. 2023 ; Zhang et al. 2023 ; Benton et al. 2024 ) . Other contributions have proposed improvements to accelerate sampling or to improve sample quality, such as ( Jolicoeur-Martineau et al. 2021 ; Bao et al. 2022 ; Watson et al. 2022 ; Xu et al. 2022 ; Rombach et al. 2022 ) . Finally, using ideas from optimal transport, ( De Bortoli et al. 2021 ; Shi et al. 2022 ; Shi et al. 2023 ) developed a parallel line of work, connecting diffusion models to Schrödinger bridge problems. 2.2.1 Learning the score for time series data Beyond image and spatial data, score-based diffusion models have also been extensively applied to time-series settings, see ( Lin et al. 2023 ) for a comprehensive survey. While the methods reviewed below are not used directly in this work, they illustrate how diffusion models have been adapted to temporal and conditional structures, and therefore provide relevant context for our setting. In the field of time series forecasting, ( Lim et al. 2023 ) draw inspiration from the TimeGrad model in ( Rasul et al. 2021 ; Kong et al. 2021 ) and the ScoreGrad model in ( Yan et al. 2021 ) to propose the TSGM model, a conditional diffusion model that operates on a low-dimensional representation of the data. ( Shen & Kwok 2023 ) remove the sequence-based architecture (LSTM, Transformer) to achieve non-autoregressive forecasting. For time-series with missing values, ( Taskhiro et al. 2021 ; Alcaraz & Strodthoff 2023 ; Kollovieh et al. 2023 ) proposed conditional diffusion frameworks that impute unobserved entries. For long-horizon multivariate sequence generation, ( Barancikova et al. 2025 ) introduce SigDiffusion, which leverages path signatures. In the domain of filtering, ( Bao et al. 2022 ; Bao et al. 2024 ) applied score-based conditional diffusion models to non-linear filtering problems. 3 Background 3.1 Score-based diffusion models Score-based diffusion models are generative models that learn a reversible transformation between two distributions, p_{0} and p_{1} . ( Song et al. 2021b ) used a forward SDE to transform samples distributed according to p_{0} to samples distributed according to p_{1} . Defining \gamma(\tau)=\gamma_{0}+\tau(\gamma_{1}-\gamma_{0}) , the forward SDE takes samples X_{0}\sim p_{0} and diffuses them towards X_{1}\sim p_{1} via a Variance-Preserving SDE (VPSDE), dX_{\tau}=-\frac{1}{2}\gamma(\tau)X_{\tau}d\tau+\sqrt{\gamma(\tau)}dW_{\tau},\ \tau\in[0,1], (6) where (W_{\tau})_{\tau\in[0,1]} is a D -dimensional Brownian motion, see Section 3.4 in ( Song et al. 2021b ) . We emphasise ( 6 ) is a different SDE to the data-generating process assumed in ( 1 ), and which defines the distribution p_{0} . From ( Anderson 1982 ) , the forward SDE in ( 6 ) admits a time-reversal which diffuses \widetilde{X}_{0}\sim p_{1} back to \widetilde{X}_{1}\sim p_{0} via, \displaystyle d\widetilde{X}_{\tau}=\left[\frac{1}{2}\widetilde{X}_{\tau}+\nabla_{\widetilde{X}_{\tau}}\ln p_{\tau}(\widetilde{X}_{\tau})\right] \displaystyle\gamma(\tau)\,d\tau (7) \displaystyle+\sqrt{\gamma(\tau)}\,d\widetilde{W}_{\tau}, such that \widetilde{X}_{\tau}\stackrel{{\scriptstyle d}}{{=}}X_{1-\tau},\widetilde{W}_{\tau}\stackrel{{\scriptstyle d}}{{=}}W_{1-\tau}\ \forall\tau\in[0,1] , where \stackrel{{\scriptstyle d}}{{=}} means equal in distribution. The reverse-time SDE in ( 7 ) requires access to the score \nabla_{X_{\tau}}\ln p_{\tau}(X_{\tau}) which is not available in closed form in general. ( Vincent 2011 ; Song et al. 2021b ) show that the oracle score-matching objective is, l(\theta)=\int_{0}^{1}\mathop{\mathbbm{E}}\Big[\lambda(\tau)\left\|s_{\theta}(\tau,X_{\tau})-\nabla_{X_{\tau}}\ln p_{\tau}(X_{\tau})\right\|_{2}^{2}\Big]d\tau, where X_{0}\sim p_{0}(x_{0}),X_{\tau}\sim q_{\tau}(x_{\tau}\mid X_{0}) and s_{\theta}:[0,1]\times\mathbbm{R}^{D}\rightarrow\mathbbm{R}^{D} is a neural network parameterised by \theta . The function \lambda:[0,1]\rightarrow\mathbbm{R}^{+} is a positive weighting function, and q_{\tau}(x_{\tau}\mid X_{0}) is the transition density of ( 6 ), q_{\tau}(x_{\tau}\mid X_{0})=\mathcal{N}(x_{\tau};\beta_{\tau}X_{0},\sigma_{\tau}^{2}I), (8) where we have defined \sigma_{\tau}^{2}=1-\beta^{2}_{\tau} and \beta_{\tau}=\exp\left[-0.25\tau^{2}(\gamma_{1}-\gamma_{0})-0.5\tau\gamma_{0}\right] . However, the objective l(\theta) cannot be evaluated in practice since we do not have access to the score. Under sufficient expressivity of the neural network s_{\theta} , ( Vincent 2011 ) show that the denoising score-matching objective has the same global minimum as l(\theta) , \displaystyle\widetilde{l}(\theta)=\int_{0}^{1}\mathbbm{E}\Big[\lambda(\tau)\big\|s_{\theta}(\tau,X_{\tau})-\nabla_{X_{\tau}}\ln q_{\tau}(X_{\tau}\mid X_{0})\big\|_{2}^{2}\Big]\,d\tau. ( Song & Ermon 2019 ) , ( Song et al. 2021b ) proposed the weighting function \lambda(\tau)=\sigma_{\tau}^{2} to reduce the variance of the target \nabla_{X_{\tau}}\ln q_{\tau}(X_{\tau}\mid X_{0})\ =-\sigma_{\tau}^{-2}(X_{\tau}-\beta_{\tau}X_{0}) . In practice, the objective \widetilde{l}(\theta) is approximated via Monte Carlo sampling. In particular, one samples (\tau,X_{0})\sim\mathcal{U}[\epsilon,1]\times p_{0} and then X_{\tau}\sim q_{\tau}(x_{\tau}\mid X_{0}) for each sample. The constant \epsilon>0 is chosen to avoid the explosion of the target since \lim_{\tau\rightarrow 0}\sigma_{\tau}=0 . 3.2 Conditional diffusion models Letting (X_{0},Y)\sim p_{0}(x_{0},y),Y\in\mathbbm{R}^{D} , ( Song et al. 2021b ) , ( Taskhiro et al. 2021 ) apply score-based diffusion models to generate samples from p_{0}(x_{0}\mid Y) . For each fixed Y , applying the forward diffusion ( 6 ) to X_{0} induces a push-forward distribution p_{\tau}(x_{\tau}\mid Y)\,\forall\,\tau\in[0,1] . Similar to ( 7 ), the reverse-time SDE for conditional generation is: \displaystyle d\widetilde{X}_{\tau}=\Big[\frac{1}{2}\widetilde{X}_{\tau}+\nabla_{\widetilde{X}_{\tau}}\ln p_{\tau}(\widetilde{X}_{\tau}\mid Y) \displaystyle\Big]\gamma(\tau)d\tau (9) \displaystyle+\sqrt{\gamma(\tau)}d\widetilde{W}_{\tau}. (Theorem 1, Appendices B.1-B.2) ( Batzolis et al. 2021 ) show one can learn the conditional score in ( 9 ) consistently by minimising the following objective, \displaystyle\tilde{L}(\theta)=\mathbbm{E}\Bigl[\lambda(\tau)\bigl\|s_{\theta}(\tau, \displaystyle X_{\tau},Y) (10) \displaystyle-\nabla_{X_{\tau}}\ln q_{\tau}(X_{\tau}\mid X_{0})\bigr\|_{2}^{2}\Bigr], The conditional score can be written in terms of the denoiser, \nabla_{X_{\tau}}\ln p_{\tau}(X_{\tau}\mid Y)=-\frac{X_{\tau}}{\sigma_{\tau}^{2}}+\frac{\beta_{\tau}}{\sigma_{\tau}^{2}}\mathbbm{E}\left[X_{0}\mid X_{\tau},Y\right], (11) where the denoiser is given by, \mathbbm{E}\left[X_{0}\mid X_{\tau},Y\right]=\int_{\mathbbm{R}^{D}}x_{0}p(x_{0}\mid X_{\tau},Y)dx_{0}. (12) To learn the denoiser directly, one can re-parameterise s_{\theta} using a network D_{\theta} , s_{\theta}(\tau,X_{\tau},Y)=-\frac{X_{\tau}}{\sigma_{\tau}^{2}}+\frac{\beta_{\tau}}{\sigma_{\tau}^{2}}D_{\theta}(\tau,X_{\tau},Y) (13) which can be plugged into the learning objective ( 10 ) for s_{\theta} to yield a learning objective for D_{\theta} , \tilde{L}(\theta)=\mathop{\mathbbm{E}}\left[\left\|D_{\theta}(\tau,X_{\tau},Y)-X_{0}\right\|_{2}^{2}\right]. (14) 4 Estimation of the drift function In this section, we study how to recover the drift \mu(Y) from a denoising model trained to approximate the denoiser \mathbb{E}[X_{0}\mid X_{\tau},Y] . We exploit the Gaussian structure of the forward process to construct a Monte Carlo estimator for \mu . 4.1 Data specification for conditional diffusion models Given trajectories Y^{(i)} observed at times t_{j}=j\Delta , recall we define \forall j\in\{0,\dots,J-1\} the increment, Z^{(i)}_{t_{j}}:=Y^{(i)}_{t_{j+1}}-Y^{(i)}_{t_{j}}. From ( 2 ), Z^{(i)}_{t_{j}} depends on the past trajectory (Y^{(i)}_{t_{s}})_{s=0}^{j} only through the current observation Y^{(i)}_{t_{j}} . Consequently, under the time-homogeneous dynamics in ( 1 ), the conditional law of Z_{t_{j}} given Y_{t_{j}} is independent of j , and we work with generic pairs (Y,Z) , suppressing the indices i and j . Henceforth, X_{0} will denote an increment from the data generating distribution p_{0}(x_{0}\mid Y) , and X_{\tau} is the diffused version of X_{0} after applying the forward process q_{\tau} in ( 8 ). We aim to learn an approximation D_{\theta}(\tau,X_{\tau},Y) to the denoising target \mathbb{E}[X_{0}\mid X_{\tau},Y] by minimising ( 10 ). 4.2 Learning the drift by denoising We suppose henceforth that \theta^{*} is the minimiser of ( 14 ) such that D_{\theta^{*}}(\tau,X_{\tau},Y) is the trained approximation to \mathbbm{E}\left[X_{0}\mid X_{\tau},Y\right] . We also use the Euler-Maruyama approximation to the increment distribution in ( 4 ), p_{0}(x_{0}\mid Y)\approx\widehat{p}_{0}(x_{0}\mid Y)=\mathcal{N}(x_{0};\mu(Y)\Delta,\Delta) to derive an estimator for \mu . Under the VPSDE, the transition kernel q_{\tau}(x_{\tau}\mid X_{0}) in ( 8 ) is a Gaussian linear in X_{0} . Since the conditional distribution \widehat{p}_{0}(x_{0}\mid Y) is also Gaussian, a closed-form solution for the denoiser is obtained under the EM approximation, \mathbbm{E}\left[X_{0}\mid X_{\tau},Y\right]\approx\frac{\sigma_{\tau}^{2}\Delta}{\sigma_{\tau}^{2}+\beta_{\tau}^{2}\Delta}\left[\mu(Y)+\frac{\beta_{\tau}}{\sigma_{\tau}^{2}}X_{\tau}\right], and the drift is obtained through, \mu(Y)=a(\tau)X_{\tau}+b_{\Delta}(\tau){\mathbbm{E}}\left[X_{0}\mid X_{\tau},Y\right], (15) where, \displaystyle a(\tau):=-\frac{\beta_{\tau}}{\sigma_{\tau}^{2}},\,\,\,\,b_{\Delta}(\tau):=\frac{\sigma_{\tau}^{2}+\beta_{\tau}^{2}\Delta}{\sigma_{\tau}^{2}\Delta}. (16) For a given Y=y , let x_{\tau}^{(k)}\sim p_{\tau}(x_{\tau}\mid y)\,\forall k\in[\mathcal{K}] . We can construct a single-sample estimator, \widehat{\mu}(\tau,x^{(k)}_{\tau},y)=a(\tau)x_{\tau}^{(k)}+b_{\Delta}(\tau)D_{\theta^{*}}(\tau,x^{(k)}_{\tau},y), (17) where x_{\tau}^{(k)} enters as a linear correction to the learned denoiser, constructed in a way that preserves unbiasedness in the ideal case of ( 15 ). By averaging over the \mathcal{K} samples x_{\tau}^{(k)} , we can estimate the drift with, \bar{\mu}(\tau,y)=\frac{1}{\mathcal{K}}\sum_{k\in[\mathcal{K}]}\widehat{\mu}(\tau,x^{(k)}_{\tau},y). (18) For sufficiently large \mathcal{K} , the estimator \bar{\mu}(\tau,y) removes the linear dependence of the rescaled denoiser on X_{\tau} in ( 15 ), and can be cast as a lower-variance alternative to \hat{\mu}(\tau,x_{\tau}^{(k)},y) . We emphasise that the Euler-Maruyama approximation used to derive ( 18 ) is not imposed during training, but only at the stage of drift estimation. 4.3 Choice of diffusion time for drift estimator To implement the estimator in ( 18 ), a value of \tau\in[0,1] must be chosen, and \mathcal{K} samples x^{(k)}_{\tau}\sim p_{\tau}(x_{\tau}\mid Y=y) must be drawn. One expects there to exist an intermediate noise level 0<\tau<1 for which the denoising problem is well conditioned and the drift approximation error in \bar{\mu} is low. We propose setting \tau=1 as a convenient and empirically robust choice, as motivated in the remainder of this section. At \tau=1 , one should first sample \mathcal{K} standard normal random variables x^{(k)}_{1}\sim\mathcal{N}(0,I) . The drift in ( 17 ) is then, \widehat{\mu}(1,x^{(k)}_{1},y)=a(1)x^{(k)}_{1}+b_{\Delta}(1)D_{\theta^{*}}(1,x^{(k)}_{1},y), Averaging over all \mathcal{K} samples yields \bar{\mu}(y):=\bar{\mu}(1,y) , \bar{\mu}(y)=\frac{1}{\mathcal{K}}\sum_{k\in[\mathcal{K}]}\widehat{\mu}(1,x^{(k)}_{1},y). (19) To justify why \tau=1 is appropriate, we empirically investigate the dependence of the estimator \bar{\mu}(\tau,y) on \tau . In principle, samples x^{(k)}_{\tau}\sim p_{\tau}(x_{\tau}\mid y) can be drawn by sampling (x_{0},y)\sim p_{0}(x_{0},y) from the dataset \mathcal{D} and applying the forward process q_{\tau} in ( 8 ). However, this approach is limited to values of y observed in the dataset and scales poorly with dimension. Therefore, sampling x^{(k)}_{\tau} from p_{\tau}(x_{\tau}\mid y) for different values of \tau is implemented by simulating the reverse diffusion ( 9 ), see ( Song et al. 2021b ; Karras et al. 2022 ) , for details on different numerical implementations. The VPSDE forward process in ( 8 ) ensures X_{\tau}\simeq\mathcal{N}(0,I) for large \tau . We can therefore sample x^{(k)}_{1}\sim\mathcal{N}(0,I) and run the reverse diffusion in ( 9 ) to obtain approximate samples from p_{\tau}(x_{\tau}\mid y) for intermediate diffusion times. During validation, we set \mathcal{K}=10 and allow oracle access to the true drift to select the model which minimises the \tau -averaged drift mean-squared error (MSE), \widehat{\theta}=\arg\min_{\theta}\mathbbm{E}_{\tau}\mathbbm{E}\left[\left\lVert\bar{\mu}(\tau,Y)-\mu(Y)\right\rVert_{2}^{2}\right], We choose to test this approach using two representative drift functions, \displaystyle\mu^{(1)}(y) \displaystyle=-y+\sin(25y),\,y\in\mathbbm{R} (20) \displaystyle\mu^{(2)}_{d}(y) \displaystyle=-4a_{d}y_{d}^{3}-2b_{d}y_{d},\,y\in\mathbbm{R}^{D},\,d\in[D] (21) where a_{d}>0,b_{d}<0\ \forall d\in[D] . Given a set of N points uniformly spaced y_{n}\in\mathcal{R}\subset\mathbbm{R}^{D} , we approximate the drift MSE as, e^{2}(\tau)\approx\frac{1}{N}\sum_{n\in[N]}\left\|\mu(y_{n})-\overline{\mu}(\tau,y_{n})\right\|_{2}^{2}. (22) Figure 1 shows the error e^{2}(\tau) when estimating \mu^{(1)} for \mathcal{K}\in\{1,10,100\} . In all cases, the drift MSE diverges as \tau\rightarrow 0 , as the approximation error in the network D_{\theta^{*}} is expected to increase for smaller \tau ( Song et al. 2021b ) . The error is minimised for \tau\approx 0.2 , but as \mathcal{K} increases, the error curve becomes flat and less sensitive to \tau , and larger values yield similar performance. Indeed, the bottom two panels in Figure 1 demonstrate accurate recovery of \mu^{(1)} for \tau\in\{0.2,1\} when \mathcal{K}=100 . Similarly, Figure 2 illustrates the drift error e^{2}(\tau) when estimating \mu^{(2)} for D=8 (top row) and D=12 (bottom row), exhibiting behaviour similar to Figure 1 even in higher dimensions. Figure 1: The top-left panel shows the drift error e^{2}(\tau) when estimating \mu^{(1)} ; the top-right panel restricts the range of \tau to \tau\in[0.15,1] . The bottom row compares the true drift with {\color[rgb]{0,0,1}\bar{\mu}}(\tau,y) for \tau=1 (left) and optimal \tau\approx 0.2 (right) for \mathcal{K}=100 . Positions Y are uniformly spaced between [-1.5,1.5] , covering 99\% of the training distribution. Figure 2: The left panels show the drift error e^{2}(\tau)\forall\tau\in[0,1] ; the right panels restrict the range of \tau to \tau\in[0.15,1] . The top row shows the drift error e^{2}(\tau) \mu^{(2)} with D=8 ; the bottom row shows the same for D=12 . We emphasise here that there is no closed form solution to selecting the optimal \tau since the approximation error in D_{\theta^{*}} is architecture and data dependent. Further, the choice of forward noise schedule will impact the dependence of the estimator on \tau , see Appendix A for a comparison with the Variance Exploding SDE. Therefore, there is no guarantee that a specific \tau is optimal across architectures and noise schedules. Moreover, the computational cost of simulating the reverse diffusion process is high. In contrast, sampling from the approximate terminal distribution X_{1}\sim\mathcal{N}(0,I) is cheap, and for large \mathcal{K} , the error at \tau=1 is similar to the optimal error. As such, we suggest using \tau=1 to implement our estimator ( 19 ), and the results in Section 5.4 confirm that even with this choice of \tau , we achieve competitive performance. 5 Experiments 5.1 Evaluation Criterion We evaluate the performance of a drift estimator \widetilde{\mu} using the time-integrated mean-squared error, E^{(\mu)}_{t}=\frac{1}{t}\int_{0}^{t}\mathbbm{E}\left\lVert\widetilde{\mu}(Y_{s})-{\mu}(Y_{s})\right\rVert_{2}^{2}ds (23) where the expectation is over the law of the process (Y_{s})_{s} . In practice, we approximate ( 23 ) in the following way. Let \{Y^{(i)}\}_{i},i\in[\mathcal{I}] be a set of i.i.d trajectories from ( 1 ) with \mathcal{J} discrete observations. We approximate ( 23 ) using \forall j\in[\mathcal{J}] , E^{(\mu)}_{j\Delta}:=\frac{1}{j\mathcal{I}}\sum_{i\in[\mathcal{I}]}\sum_{k\in[j]}\left\lVert\widetilde{\mu}(Y^{(i)}_{t_{k}})-{\mu}(Y^{(i)}_{t_{k}})\right\rVert_{2}^{2}, where \mathcal{I},\mathcal{J} denote the number of trajectories and observations, respectively, used for error evaluation. In contrast, I,J denote the number of training trajectories and observations, respectively. 5.2 Implementation In all experiments, training datasets consist of I=10^{3} sample paths generated from an Euler discretisation of ( 1 ) with \sigma=1 for t\in[0,1] at a frequency of \Delta=256^{-1} . All drift estimators are fitted with fixed random seeds for reproducibility. In-sample errors are computed by simulating \mathcal{I}=10^{3} new trajectories over the in-sample horizon t\in[0,1] with \Delta=256^{-1} , while out-of-sample errors are computed over t\in[0,20] . We will denote by \mathrm{DN} the denoising estimator \bar{\mu} in ( 19 ) which uses the trained network D_{\theta^{*}} , see Appendix B for architectural details. For one dimensional drift functions, we compare \mathrm{DN} against three nonparametric estimators used in drift estimation under i.i.d trajectories. Specifically, we implement the IID Nadaraya–Watson estimator ( Marie & Rosier 2023 ) , a kernel-based estimator obtained by averaging increments across independent trajectories; the IID Ridge ( Denis et al. 2021 ) , a spline-based least-squares estimator using B-spline bases with ridge regularization, and the IID Hermite projection ( Comte & Genon-Catalot 2020 ) , which estimates the drift via least-squares projection onto a finite-dimensional Hermite basis, see Appendix C for more details. In higher dimensions, the computational and data requirements of the Ridge and Hermite estimators of ( Denis et al. 2021 ) and ( Comte & Genon-Catalot 2020 ) grow exponentially with D ( Zhao et al. 2025 ) , and kernel-based estimators such as the Nadaraya Watson deteriorate rapidly, as seen in Appendix E . We therefore omit them from the main text when estimating high dimensional drift functions. To benchmark \mathrm{DN} in high dimensions, we implement the proposed estimator in ( Zhao et al. 2025 ) , and refer to it as \mathrm{FC} . The estimator \mathrm{FC} is the output of a feed-forward ReLU network trained to regress the scaled increments \Delta^{-1}X_{0} from the states Y using ( 5 ), corresponding to a standard regression objective, without any denoising mechanism. We follow ( Zhao et al. 2025 ) in imposing sparsity and weight-clamping constraints on the network parameters during training. These constraints restrict the hypothesis class to separable drifts and are part of the estimator definition in ( Zhao et al. 2025 ) . For consistency, all neural-network based estimators are trained to target \Delta^{-1}X_{0} . Our denoising network D_{\theta^{*}} includes a learnable embedding applied to states Y , referred to as MLPStateMapper , which combines polynomial and Fourier features and feeds into the remaining network. It also includes a parallel one-dimensional convolutional module to capture local spatial structure in (Y,X_{\tau}) , see Appendix B for more details. As the performance of \mathrm{DN} depends on both denoising-based training and architectural inductive biases in D_{\theta^{*}} , we introduce three control baselines to attempt to disentangle their contributions. We denote by \mathrm{FC}^{+} the \mathrm{FC} estimator with its parameter count matched to D_{\theta^{*}} , and remove the sparsity and weight-clamping restrictions in ( Zhao et al. 2025 ) . We further introduce \mathrm{FC}^{+} -Conv, which augments the fully connected layers with a similar parallel convolutional module as D_{\theta^{*}} . Both \mathrm{FC}^{+},\mathrm{FC}^{+} -Conv are trained on the same objective as \mathrm{FC} , i.e., the standard regression objective. Therefore, estimators \mathrm{FC}^{+},\mathrm{FC}^{+} -Conv help to isolate the effect of denoising by controlling for differences in model capacity and inductive bias. Finally, we define \mathrm{DN} -Lin as a variant of \mathrm{DN} in which the MLPStateMapper in D_{\theta^{*}} is replaced by a fully connected mapping, helping to isolate the effect of learned state embeddings in drift estimation. All estimators are selected by minimising the oracle error E^{(\mu)}_{j\Delta} at the in-sample terminal time j\Delta=1 . For the one-dimensional benchmarks, hyperparameters are chosen via grid search, see Appendix D . For neural network based estimators, ( \mathrm{FC} , \mathrm{FC}^{+} , \mathrm{FC}^{+} -Conv, \mathrm{DN} -Lin and \mathrm{DN} ), we evaluate the error E^{(\mu)}_{1} on \mathcal{I}=50 held-out trajectories at every epoch and terminate training when E^{(\mu)}_{1} has not improved over 100 epochs, see Appendix D for more details. Sensitivity to reduced training dataset size and to sampling frequency \Delta are explored in Appendix F and H , respectively, and Appendix I provides a performance comparison when we use a feasible model selection objective for \mathrm{DN},\mathrm{DN} -Lin. 5.3 Test Drift Functions For D=1 , we consider the following one-dimensional drift functions, \displaystyle\mu_{1}(y) \displaystyle=-\sin(2y)\log(1+5|y|), \displaystyle\mu_{2}(y) \displaystyle=-y+\sin(25y), \displaystyle\mu_{3}(y) \displaystyle=-y^{3}+y. To consider high-dimensional drifts derived from scalar potentials, we use the following parametric family, \displaystyle\mu_{4,d}(y)= \displaystyle-4a_{d}y_{d}^{3}-2b_{d}y_{d} \displaystyle-\frac{c}{2}\,\psi_{d}^{\prime}(y_{d})\big[\psi_{d-1}(y_{d-1})+\psi_{d+1}(y_{d+1})\big], where the coefficients satisfy a_{d}>0,b_{d}<0 , and c is a coupling constant. For c=0 , \mu_{4} collapses to a separable bi-stable potential with wells at y_{*,d}=\pm\sqrt{-\tfrac{b_{d}}{2a_{d}}} . For c>0 , coordinate interaction is introduced via \psi_{d}(y_{d})=\exp\!\big(-\tfrac{(y_{d}^{2}-y_{*,d}^{2})^{2}}{2s_{d}^{2}}\big) , and s_{d}=\sqrt{c}y_{*,d} . To consider non-potential and chaotic dynamics, we use the following drift, \forall d\in\{0,...,D-1\} and F>0 , \mu_{5,d}(y)=(y_{d+1}-y_{d-2})y_{d-1}-y_{d}+F, where indices are taken modulo D . For F<1 , \mu_{5} defines the Lorenz-96 system which exhibits stable behaviour, whereas chaotic behaviour emerges for standard choices such as F=8 ( Lorenz 1996 ) . 5.4 Results 5.4.1 In Sample Results Table 5.4.1 reports final-time in-sample errors E^{(\mu)}_{1} for the one-dimensional benchmarks. The denoising-based estimator attains errors of the same order of magnitude as classical nonparametric methods, and Figure 3 illustrates that it accurately recovers the structure of the drift. Table 1. In-sample drift errors E^{(\mu)}_{1} , for \mu_{1},\mu_{2},\mu_{3} . Best estimator is starred, second-best is underlined. E^{(\mu)}_{1} \mathrm{DN} NW Hermite Ridge \mu_{1} 0.02127 0.04564 {0.007621}^{*} 0.02382 \mu_{2} {0.01870}^{*} 0.1320 0.5049 0.0574 \mu_{3} 0.00754 0.01464 {0.005082}^{*} 0.1739 Figure 3: Panels (left to right) correspond to \mu_{1},\mu_{2},\mu_{3} . True drift against \mathrm{DN} drift. Shaded region shows 10-90\% quantile envelope over \mathcal{K} . State values cover 99\% of the training distribution. In higher dimensions, Table 5.4.1 shows that the denoising estimator \mathrm{DN} achieves the lowest in-sample error when estimating the bistable drift \mu_{4} across both separable ( c=0 ) and strongly coupled ( c=20 ) regimes. In the separable case, ( \mathrm{DN} -Lin) performs poorly in relation to \mathrm{DN} , suggesting that denoising-based drift estimation depends on an inductive bias aligned with the drift structure in order to accurately learn the drift over the in-sample distribution. Table 2. In-sample drift errors E^{(\mu)}_{1} , for \mu_{4} . Best estimator is starred, second-best is underlined. c D \mathrm{DN} \mathrm{DN} -Lin \mathrm{FC}^{+} -Conv \mathrm{FC}^{+} \mathrm{FC} 0 8 1.047 ∗ 6.941 3.898 8.865 7.687 12 1.681 ∗ 11.294 6.316 14.276 14.157 20 8 8.490 ∗ 10.335 9.048 16.287 35.365 12 10.660 ∗ 13.648 14.076 29.528 46.166 Figure 4 overlays the true and estimated drift \mathrm{DN} for selected dimensions of \mu_{4} , illustrating accurate recovery in intermediate dimensions and expected degradation in the highest-index dimension. Figure 4: True drift against the \mathrm{DN} drift for \mu_{4},c=0 . State values cover 99\% of the training distribution in each dimension. Top panels show dimensions 3 (left) and 8 (right) for D=8 . Bottom panels show dimensions 3 (left) and 12 (right) for D=12 . Shaded region denotes 10-90\% quantile envelope over \mathcal{K} . Table 5.4.1 reports in-sample results for the Lorenz 96 drift \mu_{5} . Performance is dominated by architectural inductive bias, as estimators with convolutional architectures ( \mathrm{DN} , \mathrm{DN} -Lin, \mathrm{FC}^{+} -Conv) achieve similar accuracy while fully connected baselines perform poorly. Table 3. In-sample drift errors E^{(\mu)}_{1} , for \mu_{5} . F D \mathrm{DN} \mathrm{DN} -Lin \mathrm{FC}^{+} -Conv \mathrm{FC}^{+} \mathrm{FC} \frac{1}{2} 20 4.870 3.127 ∗ 3.733 18.554 49.862 40 9.596 7.766 ∗ 11.940 44.764 62.988 8 20 9.542 5.725 ∗ 11.486 127.103 926.707 40 17.898 12.203 ∗ 16.306 1263.290 2583.505 Overall, the results indicate that denoising-based estimators yield the strongest in-sample performance when paired with a network whose inductive bias is aligned with the underlying dynamics. However, since all networks are selected by oracle in-sample error during training, these results reflect best-case fitting and may not be indicative of the relative generalisation behaviour of denoising and standard regression estimators. 5.4.2 Out of Sample Results We evaluate out-of-sample (OOS) performance by simulating trajectories beyond the training horizon, where the system in ( 1 ) evolves into regions of state space not observed during training. Table 5.4.2 reports final-time OOS errors for the bistable drift \mu_{4} . In the non-separable regime ( c=20 ), the denoising estimators ( \mathrm{DN},\mathrm{DN} -Lin) achieve the lowest OOS error compared to all standard regression baselines across dimensions. Figure 5 shows that the OOS error of \mathrm{DN} grows more slowly over time, leading to increasing separation from \mathrm{FC}^{+} -Conv. This advantage persists even when the \mathrm{FC}^{+} -Conv baseline is augmented with the MLPStateMapper state embedding (see Appendix G ), suggesting that the performance gap of \mathrm{DN} is not explained by architecture alone. Table 4. Final-time OOS drift errors E^{(\mu)}_{20} , for \mu_{4} . c D \mathrm{DN} \mathrm{DN} -Lin \mathrm{FC}^{+} -Conv \mathrm{FC}^{+} \mathrm{FC} 0 8 1.546 ∗ 11.800 9.744 14.129 14.365 12 2.314 ∗ 16.616 8.344 21.523 23.658 20 8 50.919 ∗ 67.297 76.061 70.134 67.299 12 81.721 ∗ 115.750 121.379 182.499 105.854 Figure 5: OOS E^{(\mu)}_{j\Delta} for \mu_{4},c=20 , for \mathrm{DN} , and \mathrm{FC}^{+} -Conv . Shaded region shows 10-90\% quantile envelope over test trajectories \mathcal{I} . Table 5.4.2 reports OOS results for the Lorenz 96 drift \mu_{5} . In the stable regime ( F=0.5 ), the estimators with convolutional networks achieve comparable OOS performance. In the chaotic regime ( F=8 ), the denoising-based estimator \mathrm{DN} -Lin consistently attains lower OOS error than \mathrm{FC}^{+} -Conv across dimensions. Since both estimators share the same inductive bias, these results suggest a generalisation benefit attributable to the denoising training objective rather than architectural differences. The denoising variant \mathrm{DN} underperforms in this setting, reflecting misalignment of the MLPStateMapper state embedding with the Lorenz 96 dynamics. Figure 6 shows that while OOS errors stabilise over time, \mathrm{DN} -Lin plateaus at a lower error level than standard regression. Table 5. Final-time OOS drift error E^{(\mu)}_{20} , for \mu_{5} . F D \mathrm{DN} \mathrm{DN} -Lin \mathrm{FC}^{+} -Conv \mathrm{FC}^{+} \mathrm{FC} \frac{1}{2} 20 8.718 12.024 7.920 ∗ 44.376 56.041 40 12.198 11.134 ∗ 12.802 99.322 167.826 8 20 3973 172.786 ∗ 226.361 7457 8425 40 1243 332.905 ∗ 527.726 16280 17656 Figure 6: OOS E^{(\mu)}_{j\Delta} for \mu_{5},F=8 comparing the \mathrm{DN} -Lin , \mathrm{FC}^{+} -Conv estimators. Shaded region shows 10-90\% quantile envelope across test trajectories \mathcal{I} . Taken together, the OOS results for \mu_{4} and \mu_{5} suggest standard regression, when paired with an architecture aligned the underlying drift, can achieve strong in-sample performance, yet may generalise poorly out of sample in strongly coupled or chaotic dynamics. In these settings, the denoising estimator with an appropriate architecture provides improved generalisation, while remaining competitive when networks trained via standard regression already generalise well. 6 Conclusion This paper adapts score-based conditional diffusion models to the problem of high-dimensional drift estimation from discretely observed trajectories. By introducing controlled Gaussian corruption to the observed increments, we train a neural network to learn the denoiser of each increment conditional on the previous state. Using the trained network, we derived an estimator for the drift through a first-order approximation of the data-generating process. Across a range of drift structures and dimensions, our results show that denoising can complement architectural inductive bias to improve out-of-sample robustness in settings with strong interactions or chaotic dynamics. In contrast, when standard regression already generalises stably out of sample, denoising can remain competitive. Several important questions remain open. These include developing a theoretical understanding for how different noise schedules affect the bias-variance properties of the estimator under specific architectures, as well as extending the estimator construction to incorporate higher order approximations of the denoising target. 7 Impact Statement This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here. Appendix A Sensitivity to forward noising process In this section, we motivate why the forward noising process q_{\tau}(X_{\tau}|X_{0}) needs to be carefully chosen to ensure our denoising estimator ( 19 ) accurately recovers the drift function. Recall the transition kernel for the VPSDE noising process, q_{\tau}(x_{\tau}\mid X_{0})=\mathcal{N}(x_{\tau};\beta_{\tau}X_{0},\sigma_{\tau}^{2}I), where \beta_{\tau}=\exp(-0.25\tau^{2}(\gamma_{1}-\gamma_{0})-0.5\tau\gamma_{0}) and \sigma_{\tau}^{2}=1-\beta_{\tau}^{2} . ( Song et al. 2021b ) define an alternative noising process (the Variance Exploding SDE), which takes samples X_{0}\sim p_{0} and diffuses them with the forward transition density, q^{\text{VE}}_{\tau}(x_{\tau}\mid X_{0})=\mathcal{N}(x_{\tau};X_{0},\phi_{\tau}^{2}I), (24) where \phi_{\tau}=\phi_{0}(\phi_{1}/\phi_{0})^{\tau},\forall\tau\in[\epsilon,1] . As in the VPSDE process, \epsilon>0 is chosen to avoid vanishing variance as \tau\rightarrow 0 . However, in contrast to the VPSDE process, where noise variance \sigma_{\tau}^{2} increases smoothly and saturates at 1 , the noise variance \phi_{\tau}^{2} in VESDE increases exponentially. Under uniform sampling in \tau , this concentrates most training mass at low marginal noise levels. Consequently, the denoising estimator is trained predominantly in high signal-to-noise regimes, where learning the denoising target is hardest. By comparison, the VPSDE distributes training more evenly across signal-to-noise ratios. As such, given the same maximum noise levels \phi_{1}=\sigma_{1} , the VESDE process spends more time in lower-noise regimes, where denoising is hardest. We extend the experiments in Section 4.3 by diffusing the observed increments X_{0} with the VESDE noising kernel in ( 24 ). We consider two VESDE variants. In VESDE-Matched, we set \phi_{1}=1 to match the maximum marginal variance in the VPSDE. In VESDE-Larger, we increase the maximum marginal noise to \phi_{1}=15 to compensate for the concentration of training mass at low noise levels under VESDE-Matched. In both cases, \phi_{0}=0.03 , and we set \epsilon=6.5\times 10^{-2} in VESDE-Matched and \epsilon=5\times 10^{-2} in VESDE-Larger in both training and evaluation to ensure the minimum marginal variance for both processes is identical to that of the VPSDE process in Section 4.3 . Figure 7: The top panels shows the drift error e^{2}(\tau) when estimating \mu^{(1)} ; the bottom panels compares the true drift to the estimated drift at \tau=1 . Positions Y are uniformly spaced between [-1.5,1.5] , covering 99\% of the training distribution. Figure 8: Drift error e^{2}(\tau) for \mu^{(2)} with D=8 for VESDE-Matched (left) and VESDE-Larger (right). Figure 9: Drift error e^{2}(\tau) for \mu^{(2)} with D=12 for VESDE-Matched (left) and VESDE-Larger (right). As in the VPSDE case, Figures 7 - 9 show increasing \mathcal{K} reduces the drift MSE uniformly across \tau for both VESDE variants. However, for fixed \mathcal{K} , VESDE-Matched exhibits uniformly larger errors than VESDE-Larger. This gap is most pronounced for large diffusion times, where the forward process in VESDE-Larger has higher variance. Taken together, these results show that matching maximum marginal variance is insufficient for accurate drift estimation, and that effective denoising requires adequate coverage of high noise regimes during training. Appendix B Denoising Network Architecture The architecture described in this section corresponds to the denoising network used for \mathrm{DN} . The variant \mathrm{DN} -Lin is obtained by removing the MLPStateMapper component (see Figure 11 and replacing it with a fully connected mapping of matching output dimension, and no other module is changed. The network for \mathrm{DN} is constructed such that the output is given by, D_{\theta}(\tau,X_{\tau},Y)=f_{\theta}(Y)+g_{\theta}(X_{\tau},Y)+m_{\theta}(g_{\theta},\tau). Figure 10 summarises the main components of the denoiser network architecture. The path followed by modules highlighted in blue represents the implementation of g_{\theta} as a sequence of 1D convolutional neural networks and a single 1D convolutional residual layer. The function m_{\theta} , highlighted in orange, conditions the network on the diffusion time through a fully connected residual layer. The architecture for m_{\theta} is motivated by the vanishing dependence of traditional diffusion models on the diffusion time ( Kim et al. 2024 ) . Figure 10: Score network D_{\theta} used for \mathrm{DN} . In prior diffusion model architectures, the conditioning mapping f_{\theta} is often implemented using sequence-based architectures such as Long Short-Term Memory (LSTM) networks or Transformers ( Yan et al. 2021 ; Corso et al. 2023 ; Lim et al. 2023 ; Saharia et al. 2022 ) . These architectures maintain an internal state that evolves with the history of observed inputs. In this paper, f_{\theta} is implemented using the MLPStateMapper , see Figure 11 , which defines a non-linear feature mapping of the observed state Y without maintaining a recurrent or attention-based state. This design reflects the Markovian dependence structure induced by the data-generating SDE in ( 1 ). The mapping is constructed by combining multiple feature pathways: element-wise polynomial features (Y,Y^{2},Y^{3}) , a learned global non-linear transformation implemented as a linear–ELU block, and Fourier features. The Fourier features are modulated by a learnable scalar gate, allowing their contribution to be adaptively scaled during training. All feature components are concatenated and passed through a two-layer fully connected network to produce the final output in \mathbb{R}^{D} . The number of trainable parameters in D_{\theta} scales as T(D)=2D^{2}+\mathcal{O}(D)+\mathcal{O}(1) . The quadratic dependence on the time-series dimensionality arises from the first 1D convolution in g_{\theta} , see Figure 10 , although the constant term in T(D) is much larger than the linear or quadratic terms. Figure 11: MLPStateMapper module. It provides an inductive bias towards polynomial and Fourier features, which can be learned from the data. Appendix C Baselines C.1 IID Nadaraya-Watson Estimator In ( Marie & Rosier 2023 ) , the authors improve the standard Nadaraya-Watson estimator for the drift function by averaging over \mathcal{I} IID sample paths observed discretely \mathcal{J} times. Assuming each path is defined on the same uniform time grid with \Delta=t_{j+1}-t_{j} , the discrete-time estimator for the drift function is given by, \displaystyle\widehat{b}_{\mathcal{J},\mathcal{I},h}(y)=\frac{\widehat{bf}_{\mathcal{J},\mathcal{I},h}(y)}{\widehat{f}_{\mathcal{J},\mathcal{I},h}(y)}, (25) where, \displaystyle\widehat{f}_{\mathcal{J},\mathcal{I},h}(y)=\frac{1}{\mathcal{J}\mathcal{I}}\sum_{i=1}^{\mathcal{I}}\sum_{j=0}^{\mathcal{J}-1}K_{h}(Y_{t_{j}}^{(i)}-y), (26) and, \displaystyle\widehat{bf}_{\mathcal{J},\mathcal{I},h}(y)=\frac{1}{\mathcal{I}}\sum_{i,j}K_{h}(Y_{t_{j}}^{(i)}-y)Z_{t_{j}}^{(i)}, (27) and Y_{t_{j}}^{(i)} , Z_{t_{j}}^{(i)} are defined in Section 4.1 . In our experiments, we choose K_{h}(y)=(2\pi h^{2D})^{-0.5}\exp\left[-0.5h^{-2}\left\|y\right\|_{2}^{2}\right] . To choose the optimal bandwidth h , one finds the bandwidth which minimises the leave-one-out-cross validation criterion, \displaystyle CV(h)=\sum_{i=1}^{\mathcal{I}} \displaystyle\sum_{j=0}^{\mathcal{J}-1}\Bigg[\widehat{b}^{-(i)}_{\mathcal{J},\mathcal{I},h}(Y_{t_{j}}^{(i)})^{2}\Delta \displaystyle-2\widehat{b}^{-(i)}_{\mathcal{J},\mathcal{I},h}(Y_{t_{j}}^{(i)})\big(Y_{t_{j+1}}^{(i)}-Y_{t_{j}}^{(i)}\big)\Bigg] where \widehat{b}^{-(i)}_{\mathcal{J},\mathcal{I},h} is the leave-one-out drift estimator. Under the regularity assumptions of ( Marie & Rosier 2023 ) , parameterised by a spatial smoothness index \beta\geq 1 of the marginal density q_{t}(\cdot) of ( 1 ), and a discretization parameter \epsilon\in[0,1) , risk control in expected squared L_{2} error is achieved via the truncated estimator, \displaystyle\widetilde{b}_{\mathcal{J},\mathcal{I},h}=\widehat{b}_{\mathcal{J},\mathcal{I},h}\,\mathbbm{1}\!\left(\widehat{f}_{\mathcal{J},\mathcal{I},h}>m/2\right), where m>0 is a lower bound on the time-averaged marginal density, f . Under these conditions, the estimator over a compact domain [A,B] satisfies the risk bound, \displaystyle\mathbbm{E}\bigl\|\widetilde{b}_{\mathcal{J},\mathcal{I},h}-b\bigr\|_{f,A,B} \displaystyle=\mathcal{O}\!\Big((T-t_{0})\mathcal{J}^{-1}+h^{2\beta} \displaystyle+\mathcal{I}^{-1}h^{-1}+\mathcal{I}\mathcal{J}h^{3+\epsilon}\Big). As noted in ( Marie & Rosier 2023 ) , when T>1 , choosing t_{0}\in[1,T-1] removes the time-dependent factors in the risk rate. In the worst-case regime, h<1 , the rate is maximised at \beta=1 , and if the true drift function is bounded, one may take \epsilon=0 . C.2 IID Ridge Estimator The IID Ridge estimator in ( Denis et al. 2021 ) is a nonparametric estimator of a one dimensional drift function using B -splines. Letting L_{\mathcal{I}},K_{\mathcal{I}},M>0 , and end points A_{\mathcal{I}}=-B_{\mathcal{I}}<0 , ( Denis et al. 2021 ) define a sequence of knots, \displaystyle u_{s}=\begin{cases}A_{\mathcal{I}},&s=-M,\dots,0,\\[6.0pt] A_{\mathcal{I}}+\dfrac{s}{K_{\mathcal{I}}}\bigl(B_{\mathcal{I}}-A_{\mathcal{I}}\bigr),&s=1,\dots,K_{\mathcal{I}},\\[6.0pt] B_{\mathcal{I}},&s=K_{\mathcal{I}}+1,\dots,K_{\mathcal{I}}+M.\end{cases} Denoting the spline basis by B_{s,M,u} , and the projection subspace as S_{K_{\mathcal{I}},M,{u}}=\text{span}\{(B_{s,M,{u}})\ \forall s\in\{-M,\cdots,K_{\mathcal{I}}-1\}\} , such that \text{dim}(S_{K_{\mathcal{I}},M,{u}})=K_{\mathcal{I}}+M , then for any y\in[A_{\mathcal{I}},B_{\mathcal{I}}] , the drift estimator is given by, \displaystyle\widehat{b}_{\mathcal{I},\mathcal{J}}(y)=\sum_{s=-M}^{K_{\mathcal{I}}-1}\widehat{a}_{s}B_{s,M,{u}}(y) (28) Defining the entries in the matrix {B} as B_{r,s}=B_{s,M,u}\bigl(Y^{(i)}_{t_{j}}\bigr) with row index r=\mathcal{J}(i-1)+j , and {Z}\in\mathbbm{R}^{\mathcal{I}\mathcal{J}} where Z_{r}=\Delta^{-1}(Y_{t_{j+1}}^{(i)}-Y_{t_{j}}^{(i)}) , the solution to ( 28 ) satisfies \widehat{{a}}=({B}^{T}{B})^{-1}{B}^{T}{Z} if {B}^{T}{B} is full column rank and \|\widehat{{a}}\|_{2}^{2}\leq(K_{\mathcal{I}}+M)L_{\mathcal{I}} . Alternatively, letting \widehat{\lambda} be the unique solution to \|({B}^{T}{B}+{\lambda}{I})^{-1}{B}^{T}{Z}\|_{2}^{2}=(K_{\mathcal{I}}+M)L_{\mathcal{I}} , the solution is given by \widehat{{a}}=({B}^{T}{B}+\widehat{\lambda}{I})^{-1}{B}^{T}{Z} . Assuming only the necessary conditions for the existence and uniqueness of a strong solution to ( 1 ), ( Denis et al. 2021 ) show that the risk bound is controlled by a bias term and a variance term which scales as \left(\text{dim}(S_{K_{\mathcal{I}},M,{u}})\mathcal{I}^{-1}\right)^{-0.5}+\Delta . Under further regularity conditions for the drift function, ( Denis et al. 2021 ) obtain a risk bound of order \left(\ln^{2}\mathcal{I}/\mathcal{I}\right)^{\frac{2\beta}{2\beta+1}} . C.3 IID Hermite Projection Estimator Comte and Genon-Catalot ( Comte & Genon-Catalot 2020 ) derive a nonparametric estimator for a one dimensional drift function using a least-squares projection onto a subspace spanned by Hermite polynomials. Given the first m Hermite polynomial basis functions (h_{k})_{k=0}^{m-1} , and \mathcal{I} sample paths observed continuously on t\in[0,T] , the drift estimator is given by \displaystyle\widehat{b}_{m}(y)=\sum_{k=0}^{m-1}\widehat{\theta}_{k}h_{k}(y), where (\widehat{\theta}_{k})_{k=0}^{m-1} is the vector of parameters obtained via the least-squares projection onto the space spanned by the Hermite polynomials, h_{k} , \displaystyle\widehat{{\theta}}_{k}=e_{k}^{T}\widehat{{\Phi}}_{m}^{-1}\widehat{{Z}}_{m}. (29) where e_{k} is the k -th canonical basis vector. In ( 29 ), \widehat{\Phi}_{m}\in\mathbbm{R}^{m\times m},\widehat{Z}_{m}\in\mathbbm{R}^{m} are defined as, \displaystyle\widehat{\Phi}_{m,(u,v)} \displaystyle=\frac{1}{\mathcal{I}T}\sum_{i=1}^{\mathcal{I}}\int_{0}^{T}h_{v}(Y^{(i)}_{s})h_{u}(Y^{(i)}_{s})ds \displaystyle\widehat{Z}_{m,u} \displaystyle=\frac{1}{\mathcal{I}T}\sum_{i=1}^{\mathcal{I}}\int_{0}^{T}h_{u}(Y_{s}^{(i)})dY_{s}^{(i)}, where u,v\in\{0,\cdots m-1\} . To find the optimal truncation level m , ( Comte & Genon-Catalot 2020 ) propose to optimise \displaystyle\widehat{m}=\arg_{m\in\mathcal{M}}\min\left[-\left\|\widehat{b}_{m}\right\|^{2}_{\mathcal{I}}+\kappa\left\|\widehat{{\Phi}}^{-1}_{m}\widehat{{\Phi}}_{m,\sigma^{2}}\right\|_{op}\frac{m}{\mathcal{I}T}\right], where \mathcal{M}=\left\{m:m\leq 10,m\left\|\widehat{{\Phi}}^{-1}_{m}\right\|^{1/4}_{op}\leq\mathcal{I}T\right\} , see ( Comte & Genon-Catalot 2020 ) for notation. In the case of continuous observations of the sample paths and bounded diffusion coefficient, Theorem 1 in ( Comte & Genon-Catalot 2020 ) proves the risk bound on \widehat{b}_{\widehat{m}} is order \mathcal{O}(1/\mathcal{I}T) . C.4 Neural-Network Based Benchmark Estimators The benchmark \mathrm{FC} is constructed as a fully-connected feedforward network with ReLU activation functions, as in ( Zhao et al. 2025 ) . The number of layers and the width of each layer are chosen based on the empirical results in the code repository by ( Zhao et al. 2025 ) . However, for a fixed dataset, the number of parameters in \mathrm{FC} can be several orders of magnitude smaller than our denoising network D_{\theta} . The benchmark \mathrm{FC}^{+} increases the depth of \mathrm{FC} until the number of parameters exceeds the parameter count in D_{\theta} . Architectural details achieving this constraint are fixed deterministically and reported in the code repository. We implement \mathrm{FC^{+}} -Conv as the sum of a fully-connected backbone and a convolutional module, (\mathrm{FC}^{+}\text{-Conv})_{\theta}(Y)\equiv\mathrm{FC}_{\theta}(Y)+g_{\theta}(Y), The convolutional module g_{\theta} follows the same design pattern as the convolutional path in \mathrm{DN} , but with the first channel dimension adjusted to account for the absence of noisy inputs X_{\tau} . The fully-connected backbone of \mathrm{FC^{+}} -Conv is resized relative to \mathrm{FC^{+}} so that the total number of trainable parameters of \mathrm{FC^{+}} -Conv matches that of D_{\theta} for the same dataset. Appendix D Estimator hyperparameter selection D.1 IID Hermite Estimator Figure 12 shows how the number of basis functions for the Hermite estimator in ( Comte & Genon-Catalot 2020 ) is selected by optimising E^{(\mu)}_{1} . As suggested in ( Comte & Genon-Catalot 2020 ) , increasing the number of basis functions beyond m=10 yields small marginal improvements in estimator accuracy; nevertheless, we conduct a grid search for m>10 for completeness. Figure 12: Hermite drift error E^{(\mu)}_{1} , against number of basis functions. From left to right, top to bottom, the estimation for \mu_{1} , \mu_{2} , \mu_{3} . D.2 IID Ridge Estimator Figure 13 shows how the number of subspace dimensions for the Ridge estimator from ( Denis et al. 2021 ) is chosen by minimising E^{(\mu)}_{1} . The Ridge estimator requires a higher B -spline dimension to accurately estimate \mu_{2} , plausibly due to the presence of a high-frequency sinusoidal component in the drift function. In contrast, for \mu_{1},\mu_{3} , a relatively small number of B -splines are sufficient for accurate drift recovery. Figure 13: Ridge drift error E^{(\mu)}_{1} , against number of basis functions. From left to right, top to bottom, the estimation for \mu_{1} , \mu_{2} , \mu_{3} . D.3 IID Nadaraya Estimator Figures 14 and 15 illustrate how the bandwidth is chosen for the Nadaraya-Watson estimator when estimating \mu_{4} and \mu_{5} . For small bandwidths, the Nadaraya estimator discards many training samples, leading to large errors. As the bandwidth increases, the error decreases until it rises again once the estimator enters an over-diffuse regime in which all training points are weighted uniformly. Figure 15 shows that the optimal bandwidth increases slightly with dimension, as expected. Figure 14: Nadaraya drift error E^{(\mu)}_{1} versus bandwidth h for estimation of \mu_{4} . Figure 15: Nadaraya drift error E^{(\mu)}_{1} versus bandwidth h for estimation of \mu_{5} . Top panels show D=20,F\in\{0.5,8\} (left to right); bottom panels show D=40,F\in\{0.5,8\} . D.4 Denoising Estimator We train the network D_{\theta} with the MLPStateMapper module from Figure 11 using Adam ( Kingma & Ba 2015 ) with an initial learning rate of 10^{-2} . We allow the learning rate to decrease by a constant factor of 0.9 whenever the training loss fails to improve for 60 consecutive epochs, setting the minimum learning rate at 10^{-3} . The output weights in g_{\theta},m_{\theta} are initialised to zero. For the conditional diffusion model hyperparameters, we set \epsilon=10^{-3},S=10^{4},\gamma_{0}=0,\gamma_{1}=20 . We train on a single NVIDIA GeForce RTX 3090 GPU (24 GiB of memory), using Pytorch 2.7 and CUDA 12.4 with driver 550.54.15. All denoising networks are trained until the validation error E_{256}^{(\mu)} has stopped improving for 100 epochs, as illustrated in Figure 16 for a selection of drift classes. In red, we show the value of E^{(\mu)}_{1} evaluated at each epoch, suggesting the denoising objective closely matches the learning error in the drift. Figure 16: Training loss and {\color[rgb]{1,0,0}E^{(\mu)}_{1}} loss for \mathrm{DN} (left panels), and \mathrm{DN} -Lin (right panels). Left scale shows training loss, right scale shows validation loss. Figure 17 shows the batch-averaged norm of the gradients of the network output D_{\theta}(\tau,X_{\tau},Y) during training of \mathrm{DN} and \mathrm{DN} -Lin. Across the selected drift classes, we observe that the gradients with respect to each of the inputs do not vanish across epochs. Figure 18 shows the batch-averaged ratio \left\lVert m_{\theta}\right\rVert/\left\lVert D_{\theta}\right\rVert generally increases during training, suggesting the contribution of the diffusion time embedding grows before the model selection occurs. Taken together, Figures 17 and 18 confirm that the architecture in Appendix B does not collapse to standard regression using Y . Figure 17: Training-time gradient norms with respect to diffusion time , {\color[rgb]{0.1216,0.4667,0.7059}X_{\tau}} , and {\color[rgb]{1,0.5,0}Y} . Left panels show \mathrm{DN} , right panels show \mathrm{DN} -Lin. Figure 18: Relative norm of m_{\theta} to network output D_{\theta} during training. Left panels show \mathrm{DN} , right panels show \mathrm{DN} -Lin. Vertical line indicates model selection epoch. Appendix E Kernel Drift Estimators in High Dimensions For completeness, Table E shows the out-of-sample performance of the Nadaraya-Watson estimator from ( Marie & Rosier 2023 ) . As highlighted in Section 5.2 and Appendix C , the kernel bandwidth is chosen through a grid search to minimise E_{1}^{(\mu)} . Table E shows that the Nadaraya–Watson estimator performs poorly in high dimensions when compared to all baselines reported in Section 5.4.2 . Table 6. Final-time out-of-sample drift error E^{(\mu)}_{1} for kernel (NW) estimators. c D \mathrm{NW} 0 8 14.454 12 26.544 20 8 292.256 12 231.207 (a) \mu_{4} F D \mathrm{NW} \tfrac{1}{2} 20 186.914 40 220.199 8 20 8493 40 17237 (b) \mu_{5} Appendix F Sensitivity to number of training trajectories We quantify the sensitivity of the final-time in-sample error E_{1}^{(\mu)} to a tenfold reduction in the number of training paths, from I=10^{3} to I=10^{2} . Training procedure is identical to that outlined in Section 5.2 , and we report the in-sample drift error E^{(\mu)}_{1} for each test drift from Section 5.3 . We define the relative change in the error as \left|E^{(\mu)}_{1}(10^{2})-E^{(\mu)}_{1}(10^{3})\right|/E^{(\mu)}_{1}(10^{3}) . Tables F - F summarize the results for the estimators (\mathrm{DN}) , (\mathrm{DN} -Lin), and \mathrm{FC}^{+} -Conv, for the bistable potential ( \mu_{4} ) and Lorenz 96 ( \mu_{5} ) drift families. We exclude the \mathrm{FC} and \mathrm{FC}^{+} baselines from this section, as they are not competitive in the corresponding in-sample regimes in Section 5.4 . Table 7. In-sample drift error E^{(\mu)}_{1} of \mathrm{DN} with 10^{2} training paths, for \mu_{4} and \mu_{5} . Relative change is computed with respect to the corresponding 10^{3} -path results reported in Section 5.4 . Drift Param D E^{(\mu)}_{1}(10^{2}) Rel. change \mu_{4} c=0 8 5.483 4.24 12 10.415 5.20 c=20 8 17.188 1.03 12 19.638 0.84 \mu_{5} F=0.5 20 22.629 3.65 40 38.291 2.99 F=8 20 60.648 5.36 40 115.197 5.44 Table 8. In-sample drift error E^{(\mu)}_{1} of \mathrm{DN} -Lin with 10^{2} training paths, for \mu_{4} and \mu_{5} . Drift Param D E^{(\mu)}_{1}(10^{2}) Rel. change \mu_{4} c=0 8 12.517 0.80 12 19.913 0.76 c=20 8 11.935 0.16 12 15.483 0.13 \mu_{5} F=0.5 20 19.172 5.13 40 28.209 2.63 F=8 20 53.942 8.42 40 85.047 5.97 Table 9. In-sample drift error E^{(\mu)}_{1} of \mathrm{FC}^{+} -Conv with 10^{2} training paths, for \mu_{4} and \mu_{5} . Drift Param D E^{(\mu)}_{1}(10^{2}) Rel. change \mu_{4} c=0 8 11.425 1.93 12 19.065 2.02 c=20 8 21.307 1.36 12 45.019 2.20 \mu_{5} F=0.5 20 18.383 3.93 40 30.650 1.57 F=8 20 47.310 3.12 40 94.261 4.78 Across all settings, reducing the number of training paths from I=10^{3} to I=10^{2} leads to a consistent increase in final-time in-sample error. While the relative ordering of the estimators is not preserved relative to Section 5.4 , the observed changes reflect differences in the magnitude of degradation, rather than the collapse of any particular estimator, as no estimator exhibits abrupt divergence behaviour under reduced data. Appendix G Extended Out-of-Sample Results We report additional results for the bistable drift \mu_{4} , and we extend the main comparison in Section 5.4.2 to a higher dimension ( D=20 ) for the case c=20 . Tables G and G include the denoising estimator ( \mathrm{DN} ) and the convolutional baseline ( \mathrm{FC}^{+} -Conv). We additionally introduce ( \mathrm{FC}^{+} -Conv-MLPSM), in which the fully connected layers in \mathrm{FC}^{+} -Conv are replaced by the MLPStateMapper in \mathrm{DN} . Training and evaluation procedures are identical to those reported in Section 5.2 . Table 10. Final-time OOS drift error E^{(\mu)}_{20} , for \mu_{4},c=0 . c D \mathrm{DN} \mathrm{FC}^{+} -Conv-MLPSM \mathrm{FC}^{+} -Conv 0 8 1.546 0.793 ∗ 9.744 12 2.314 1.695 ∗ 8.344 Table 11. Final-time OOS drift error E^{(\mu)}_{20} , for \mu_{4},c=20 . c D \mathrm{DN} \mathrm{FC}^{+} -Conv-MLPSM \mathrm{FC}^{+} -Conv 20 8 50.919 ∗ 124.972 76.061 12 81.721 ∗ 111.634 121.379 20 182.566 ∗ 278.475 245.536 In the separable regime (c=0) , the MLPStateMapper leads to improved performance of standard regression baselines, as ( \mathrm{FC}^{+} -Conv-MLPSM) lower OOS errors than its denoising-based counterpart. In contrast, in the non-separable regime (c=20) , \mathrm{FC}^{+} -Conv-MLPSM consistently underperforms \mathrm{DN} across all dimensions, indicating that architectural bias alone is insufficient to recover the observed OOS gains. Appendix H Sensitivity to time series sampling frequency As motivated in the Introduction, in increment-based regression, decreasing the sampling frequency \Delta can lead to high variance estimators. In this section, we study the stability of the denoising estimator to different sampling frequencies \Delta and compare it to the estimator \mathrm{FC}^{+} -Conv-MLPSM (see Appendix G ). We consider the bi-stable drift \mu_{4} in both separable (c=0) and non-separable (c=20) regimes. We exclude the Lorenz 96 system from this analysis, as its behaviour is dominated by advective dynamics, making a direct comparison of sampling-frequency effects less informative. Table 12. Final time out-of-sample error E_{20}^{(\mu)} for \mu_{4} and \mathrm{DN} and \mathrm{FC}^{+} -Conv-MLPSM for different sampling frequencies \Delta . c \Delta \mathrm{DN} \mathrm{FC}^{+} -Conv-MLPSM 0 64^{-1} 0.953 0.784 256^{-1} 1.546 0.793 1024^{-1} 1.582 0.764 20 64^{-1} 119.032 92.415 256^{-1} 50.919 124.972 1024^{-1} 90.617 162.930 Table H shows that out-of-sample performance depends on the sampling frequency \Delta for both estimators, exhibiting non-monotone behaviour as \Delta decreases. In the non-separable regime, \mathrm{FC}^{+} -Conv-MLPSM performs better at coarser sampling ( \Delta=1/64 ) than at finer sampling ( \Delta=1/1024 ), while the denoising estimator \mathrm{DN} remains superior as \Delta is reduced. Appendix I Feasible Model Validation We consider a feasible model validation strategy for denoising-based estimators \mathrm{DN} , \mathrm{DN} -Lin. The neural network-based baselines in Section 5.4 are trained on ( 14 ) and the best model is selected by early stopping based on the drift error E_{1}^{(\mu)} . Since this requires knowledge of the true drift, we propose an alternative that does not require oracle access to the drift. In particular, we define the validation error using the training objective ( 14 ), evaluated on \mathcal{I}=50 held out paths. We report the out-of-sample performance of \mathrm{DN},\mathrm{DN} -Lin using oracle and feasible validation strategies in Tables I - I below. Table 13. Final-time out-of-sample drift error E_{20}^{(\mu)} , for \mathrm{DN} under oracle and feasible validation. Drift Param D Oracle Feasible \mu_{4} c=0 8 1.546 3.361 12 2.314 3.290 c=20 8 50.919 50.218 12 81.721 78.356 \mu_{5} F=0.5 20 8.718 12.105 40 12.198 10.680 F=8 20 3973.37 1854.08 40 1243.11 1356.77 Table 14. Final-time out-of-sample drift error E_{20}^{(\mu)} , for \mathrm{DN} -Lin under oracle and feasible validation. Drift Param D Oracle Feasible \mu_{4} c=0 8 11.800 18.327 12 16.616 25.849 c=20 8 67.297 64.894 12 115.750 109.155 \mu_{5} F=0.5 20 12.024 10.008 40 11.134 14.835 F=8 20 172.783 346.931 40 332.905 371.282 Across the settings considered, Tables I and I show feasible validation selects models of the same order of performance as oracle selection for the same estimator, without inducing catastrophic outcomes. Occasional cases where feasible validation yields an estimator with lower out-of-sample error than oracle selection are expected, since the oracle criterion targets drift recovery under the in-sample distribution, which can be significantly different from the out-of-sample distribution. These results indicate that feasible validation can be sufficient for model selection in practice, without relying on access to the true drift.