Organizations: Department of Electrical and Computer Engineering, University of Illinois Urbana-Champaign · Marshall School of Business, University of Southern California · Department of Computer Science, Stanford University
Complex systems often evolve in continuous time but are observed only at discrete time points with measurement noise. When the interactions can change across unobserved states, recovering both the state transitions and the differential equations is important for mechanistic insight and for identifying time-varying directed graphs. However, this recovery is challenging in high dimensions. In this paper, we study estimation and graph recovery for high-dimensional Markov-switching ordinary differential equation models from noisy discrete observations. We model the switching between states by a latent continuous-time Markov chain, and within each state the drift is a sparse additive function with unknown nonlinear components. We approximate each component by a finite basis expansion and estimate the parameters under group sparsity. Our procedure estimates the latent trajectories by wavelet smoothing and then runs a penalized surrogate expectation-maximization algorithm. Under stationarity and geometric β-mixing, given a sufficiently accurate estimated trajectory and other local regularity conditions, we show that the algorithm converges geometrically to a neighborhood of the true parameters. Finally, in an application to resting-state fMRI data, the procedure identifies multiple connectivity states, whose occupation times we compare descriptively between the children with combined-type ADHD and the typically developing children used in the fit.
Figures & tables
Figure 1: Pooled ROC curves for the two data-generating processes over 10 runs (independent measurement-noise replicates conditional on one fixed latent path), using FPR bins of width 0.1 as described in Appendix \thechapter.L.2 , with X(t) estimated by the wavelet smoothing of Section 3.1 . The area under a pooled curve can differ substantially from the mean AUC in Table 1 . Each panel plots the true positive rate (y-axis) against the false positive rate (x-axis). Top row. Results of the proposed method. The AUC generally improves with increasing sample size. Bottom row. Results of the oracle method, where the latent Markov chain Z(t) is assumed to be known. The oracle method typically achieves a higher AUC at the same sample size.
Data-Generating Process 1
Data-Generating Process 2
Proposed Method
Oracle Method
Proposed Method
Oracle Method
State
1
2
1
2
1
2
1
2
N
40
—
—
0.89(0.025)
0.88(0.020)
0.53(0.025)
0.60(0.016)
0.77(0.021)
0.50(0.037)
80
0.53(0.048)
0.52(0.073)
0.90(0.025)
0.90(0.039)
0.62(0.068)
0.83(0.027)
0.79(0.026)
0.92(0.012)
120
0.63(0.173)
0.64(0.177)
0.91(0.017)
0.90(0.023)
0.81(0.128)
0.95(0.005)
0.93(0.018)
1.00(0.002)
Table 1: The AUCs of the two data-generating processes in Figure 1 . Each numerical cell gives the mean AUC and, in parentheses, its standard deviation over all 10 runs, which are independent measurement-noise replicates conditional on one fixed latent path. A dash indicates that some runs have no valid ROC curve under the state-matching rule, defined in Appendix \thechapter.L.2 . The oracle method assumes that the latent Markov chain is known.
Figure 2: Panels (a)–(c) show the estimated graphs of states 1 – 3 as connectomes on the 116 AAL regions. Appendix \thechapter.N.4 gives the full summaries. Red and blue arrows indicate positive and negative estimates of θijℓ ; darker colors indicate larger magnitudes within each panel. Panel (d) shows the estimated transition rate matrices. Panel (e) shows the estimated smoothing probabilities P(Z(tn)=ℓ∣Y0N;Θgroup) for each subject.
group
state 1
state 2
state 3
total occupation time
TDC
2216.10
1960.19
953.71
ADHD-C
2802.16
1338.80
989.04
average occupation time ( τℓ,group )
TDC
147.74(30.068)
130.68(36.884)
63.58(31.534)
ADHD-C
186.81(48.278)
89.25(32.223)
65.94(21.425)
Table 2: Posterior expected occupation time of each state by group, in seconds, computed with a sampling period of h=2.0 seconds. The rows labeled “total occupation time” sum over the 15 subjects of each group, and the rows labeled “average occupation time” give τℓ,group , with the standard deviation across the 15 subjects in parentheses.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Two Hidden Markov Models
Figure 4: The graph associated with state ℓ=1 has self-loops on nodes X1 and X2 , and bidirectional edges between nodes X1 and X2 , X3 and X4 , and X5 and X6 .
Figure 5: The graph associated with state ℓ=2 has self-loops on nodes X5 and X6 , and bidirectional edges between nodes X5 and X6 , X7 and X8 , and X9 and X10 .
Figure 6: The simulated trajectories of data-generating process 1, defined in ( 9 )–( 10 ), with the parameters of Section \thechapter.L.3 . The blue solid lines denote Xi(t) for i=1,…,10 , the black dots denote the noisy observations Yn,i for i=1,…,10 and n=0,…,N−1 , and the grey dashed lines mark the times at which the latent Markov chain Z(t) switches states.
Figure 7: The graph of state 1 of the data-generating process 2 .
Figure 8: The graph of state 2 of the data-generating process 2 .
Figure 9: The simulated trajectories of data-generating process 2, defined in ( 11 )–( 12 ), with the parameters of Section 5 . We plot the first half [0,20] of the time interval [0,40] and the first 10 of the p=20 nodes. The blue solid lines denote Xi(t) , the black dots denote the noisy observations Yn,i for i=1,…,10 at the observation times in [0,20] , and the grey dashed lines mark the times at which the latent Markov chain Z(t) switches states.
Figure 10: Box plots of the number of basis functions m and the penalty parameter λ selected at k=2 by the procedure described in Appendix \thechapter.L.1.3 . The orange line denotes the median, the upper line of the box denotes the third quartile (Q 3 ), and the lower line of the box denotes the first quartile (Q 1 ). The whiskers extend from the box toward Q 3+1.5 IQR and Q 1−1.5 IQR, where IQR is the inter-quartile range. The dots denote the outliers that lie outside the ends of the whiskers. The true number of states is 2 for both cases, and the true number of basis functions is 3 for process 1 and 1 for process 2 . When the sample size is sufficiently large, the selected number of basis functions equals the true one in most runs ( m=3 for process 1 and m=1 for process 2 ). Furthermore, for process 1 the selected λ decreases as the sample size increases, and for process 2 it tends to decrease for N≥120 .
Figure 11: The distances of the estimated parameters to the true parameters for N≤200 , at the true k⋆=2 and m⋆ and with only λ selected by the procedure described in Appendix \thechapter.L.1.3 . The top row shows process 1 , where the distances drop between N=120 and N=160 , sharply for σ2 and modestly for {θijℓ} ; at N≤120 , the median distance of {θijℓ} is that of the zero estimate. The bottom row shows process 2 , where the distances of {θijℓ} and σ2 decrease as the sample size increases.
Figure 12: Pooled ROC curves of the two data-generating processes over 10 runs (independent measurement-noise replicates conditional on one fixed latent path), with the true X(t) given and FPR bins of width 0.1 , using the pooling procedure described in Appendix \thechapter.L.2 . The area under a pooled curve can differ substantially from the mean AUC in Table 2 . Top row. Results of the proposed method. Bottom row. Results of the oracle method, where the latent Markov chain Z(t) is assumed to be known; its fits are evaluated with the identity permutation at every λ . The variation across runs and the comparison with estimated trajectories are discussed in the text.
Data-Generating Process 1
Data-Generating Process 2
Proposed Method
Oracle Method
Proposed Method
Oracle Method
State
1
2
1
2
1
2
1
2
N
40
—
—
0.90(0.016)
0.80(0.027)
0.53(0.049)
0.84(0.010)
0.86(0.006)
0.90(0.005)
80
0.93(0.056)
0.83(0.029)
0.94(0.026)
0.92(0.012)
0.81(0.059)
0.98(0.002)
1.00(0.001)
1.00(0.000)
120
0.92(0.046)
0.85(0.040)
0.95(0.017)
0.93(0.014)
0.90(0.135)
0.98(0.000)
1.00(0.000)
1.00(0.000)
Appendix
Table 2: The AUCs of the two data-generating processes in Figure 12 , with the true X(t) given. Each numerical cell gives the mean AUC and, in parentheses, its standard deviation over all 10 runs; the runs are independent measurement-noise replicates conditional on one fixed latent path: they share Z(t) , X(0) and X(t) and differ only in the measurement noise. A dash indicates that some runs have no valid ROC curve under the state-matching rule: this occurs in 2 of the 10 runs for the proposed method on process 1 at N=40 , as reported in Appendix \thechapter.L.2 . The oracle method additionally assumes that the latent Markov chain is known and is evaluated with the identity permutation at every λ .
Figure 13: Pooled ROC curves of the two data-generating processes over 10 runs (independent measurement-noise replicates conditional on one fixed latent path), with FPR bins of width 0.1 , using the pooling procedure described in Appendix \thechapter.L.2 . The area under a pooled curve can differ substantially from the mean AUC in Table 3 . The orange, green, red and purple curves correspond to N=80 , 120 , 160 and 200 , respectively, as in Figure 1 . Top row. Results of the MS-VAR baseline. Its posterior concentrates on one state as N grows, as discussed in Appendix \thechapter.M , and state 1 of data-generating process 1 is not recovered at any sample size. Bottom row. Results of the oracle MS-VAR baseline, where the sampled chain {Z(tn)} is assumed to be known; its fits are evaluated with the identity permutation at every λ .
Data-Generating Process 1
Data-Generating Process 2
MS-VAR Baseline
Oracle MS-VAR Baseline
MS-VAR Baseline
Oracle MS-VAR Baseline
State
1
2
1
2
1
2
1
2
N
80
0.65(0.070)
0.82(0.020)
0.83(0.007)
0.64(0.030)
0.78(0.048)
0.86(0.014)
0.82(0.009)
0.94(0.002)
120
0.58(0.030)
0.79(0.025)
0.84(0.003)
0.67(0.012)
0.85(0.041)
0.91(0.003)
0.93(0.006)
0.98(0.001)
160
0.62(0.047)
0.81(0.025)
0.84(0.004)
0.72(0.014)
0.72(0.077)
0.91(0.002)
0.94(0.003)
0.96(0.001)
Appendix
Table 3: The AUCs of the two data-generating processes in Figure 13 . Each numerical cell gives the mean AUC and, in parentheses, its standard deviation over all 10 runs. The runs are independent measurement-noise replicates conditional on one fixed latent path: they share Z(t) , X(0) and X(t) and differ only in the measurement noise. The states of the MS-VAR baseline are matched as described in Appendix \thechapter.M . The oracle MS-VAR baseline assumes that the latent Markov chain is known and is evaluated with the identity permutation.
State 1
State 2
State 3
Nonzero estimates (density)
7075(53.0%)
4961(37.2%)
1873(14.0%)
Positive / negative nonzero estimates
3325/3750
2387/2574
879/994
Total absolute magnitude
66.88
52.69
14.82
Mean absolute magnitude
0.0095
0.0106
0.0079
Edges displayed
134
133
134
Appendix
Table 4: Connectivity summaries for the estimates of θijℓ used in Figure 2 (a)–(c). All summaries exclude diagonal entries; there are 116×115=13,340 possible directed edges. Density is the proportion of nonzero off-diagonal estimates, and the mean absolute magnitude is computed over those nonzero estimates.
Causal representation learning for time series has developed strong identifiability results in discrete-time latent causal models, but identifiability in continuous-time latent stochastic differential equation (SDE) models remains largely open. We address this gap using environment-induced shifts in diffusion covariance. We study additive-noise latent SDEs observed through an unknown nonlinear diffeomorphism, with shared drift but environment-specific diffusion covariance. We show that two diagonal diffusion regimes with pairwise distinct coordinate-wise variance ratios identify the latent coordinates up to permutation, coordinate-wise scaling, and a possible constant shift, without any sparsity assumption on the drift. We first prove this result for linear Ornstein-Uhlenbeck systems and then extend it to general additive-noise latent SDEs. Under mild smoothness, the instantaneous drift-Jacobian causal graph is identifiable up to the same permutation. We propose a two-stage estimator for latent disentanglement and optional graph recovery; experiments on synthetic systems confirm the predicted identifiability boundary, and an application to Hardanger Bridge monitoring data illustrates the approach on real sensor trajectories.
Yuanyuan Wang, Wenjie Wang, Haoxuan Li +2
Mohamed bin Zayed University of Artificial Intelligence · The University of Melbourne · Peking University +1
Stochastic differential equations (SDEs) are a fundamental tool for modelling dynamic processes, including gene regulatory networks (GRNs), contaminant transport, financial markets, and image generation. However, learning the underlying SDE from data is a challenging task, especially if individual trajectories are not observable. Motivated by burgeoning research in single-cell datasets, we present the first comprehensive approach for jointly identifying the drift and diffusion of an SDE from its temporal marginals. Assuming linear drift and additive diffusion, we show that non-identifiability can only arise if the initial distribution possesses generalized rotational symmetries. We further prove that even if this condition holds, the drift and diffusion can almost always be recovered from the marginals. Additionally, we show that the causal graph of any SDE with additive diffusion can be recovered from the identified SDE parameters. To complement this theory, we adapt entropy-regularized optimal transport to handle anisotropic diffusion and introduce APPEX (Alternating Projection Parameter Estimation from X0), an iterative algorithm designed to estimate the drift, diffusion, and causal graph of an additive noise SDE, solely from temporal marginals, and demonstrate its effectiveness on simulated data from linear additive noise SDEs.
Vincent Guan, Joseph Janssen, Hossein Rahmani +4
Department of Mathematics University of British Columbia, Vancouver, Canada · Department of Earth, Ocean and Atmospheric Sciences University of British Columbia, Vancouver, Canada · School of Mathematics and Statistics University of Melbourne, Melbourne, Australia
We develop a semi-parametric state-space model for time-series data with latent regime transitions. Classical Markov-switching models use fixed parametric transition functions, such as logistic or probit links, which restrict flexibility when transitions depend on nonlinear and context-dependent effects. We replace this assumption with learned functions f0,f1∈\calH, where \calH is either a reproducing kernel Hilbert space or a spline approximation space, and define transition probabilities as pjk,t=\sigmoid(f(\bxt−1)). The transition functions are estimated jointly with emission parameters using a generalized Expectation-Maximization algorithm. The E-step uses the standard forward-backward recursion, while the M-step reduces to a penalized regression problem with weights from smoothed occupation measures. We establish identifiability conditions and provide a consistency argument for the resulting estimators. Experiments on synthetic data show improved recovery of nonlinear transition dynamics compared to parametric baselines. An empirical study on financial time series demonstrates improved regime classification and earlier detection of transition events.
Prakul Sunil Hiremath
Visvesvaraya Technological University, Belagavi, India · Aliens on Earth (AoE) Autonomous Research Group, Belagavi, India