Cascaded-IO: Robust State Estimation

• Cascaded-IO is a robust filtering technique in state-space models that employs Huber-type clipping to counteract innovation outliers.
• It adapts the classical Kalman filter by cascading an outlier detection and correction step to efficiently track regime shifts.
• The method achieves minimax MSE optimality by dynamically adjusting filter gains based on a predefined contamination radius.

Dedicated-IO, in the context of robust state estimation, refers to Kalman filter procedures tailored to address innovation outliers (IOs)—i.e., large, possibly adversarial deviations within the process noise channel of a linear Gaussian state-space model. These models arise in time-discrete systems described by equations of the form $X_t = F_t X_{t-1} + v_t$ and $Y_t = Z_t X_t + \varepsilon_t$, where IOs constitute endogenous or propagating disturbances affecting the state evolution. The dedicated IO-robust approach is concretely instantiated by the rLS.IO filter, which is constructed to be minimax mean-squared-error (MSE) optimal under a specified radius of IO contamination by employing nonlinear, piecewise-linear “Huber-type” corrections to the standard Kalman filter update mechanism. This approach provides a mathematically principled and computationally efficient means to aggressively track regime changes in the system state, outperforming both the classical Kalman filter and purely additive-outlier (AO)-robust variants in the relevant contamination setting (Ruckdeschel, 2010).

1. State-Space Model and Classical Kalman Framework

The dedicated IO-robust filter assumes a time-discrete linear-Gaussian state-space model: $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$ with known matrices $F_t$, $Z_t$, i.i.d. Gaussian noise terms $v_t \sim N_p(0, Q_t)$, $\varepsilon_t \sim N_q(0, V_t)$, and mutually independent initial and noise variables. The classical (non-robust) Kalman filter proceeds by prediction: $$\begin{align*} \hat{X}_{t|t-1} &= F_t \hat{X}_{t-1|t-1}, \ \Sigma_{t|t-1} &= F_t \Sigma_{t-1|t-1} F_t^\top + Q_t, \end{align*}$$ and correction: $$\begin{align*} \Delta Y_t &= Y_t - Z_t \hat{X}_{t|t-1}, \ M^0_t &= \Sigma_{t|t-1} Z_t^\top \left( Z_t \Sigma_{t|t-1} Z_t^\top + V_t \right)^{-1}, \ \hat{X}_{t|t} &= \hat{X}_{t|t-1} + M^0_t \Delta Y_t. \end{align*}$$ This framework, however, is not robust to gross deviations (outliers) in the process noise.

2. IO-Contamination Model and Minimax Robustness

Innovation outliers are modeled by perturbing the conditional law of $X_t$ given $Y_t = Z_t X_t + \varepsilon_t$0 away from the nominal $Y_t = Z_t X_t + \varepsilon_t$1, introducing a contamination neighborhood: $Y_t = Z_t X_t + \varepsilon_t$2 where at each time $Y_t = Z_t X_t + \varepsilon_t$3, with probability $Y_t = Z_t X_t + \varepsilon_t$4 the system evolves as in the ideal model, and with probability $Y_t = Z_t X_t + \varepsilon_t$5, the innovation is drawn from an arbitrary (unspecified) distribution.

The robust estimation goal is the minimax-MSE: $Y_t = Z_t X_t + \varepsilon_t$6 thereby targeting uniformly good performance against a worst-case IO contamination of radius $Y_t = Z_t X_t + \varepsilon_t$7.

3. rLS.IO Filter Algorithm and Huberization

The dedicated IO-robust filter, rLS.IO, optimally addresses IOs via a specific structure. The estimation is based on the “one-step” innovation: $Y_t = Z_t X_t + \varepsilon_t$8 and the observed residual: $Y_t = Z_t X_t + \varepsilon_t$9

Classically, the best linear estimator is $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$0, but with IOs, large excursions in $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$1 must not be “shrunk” towards prior estimates. The rLS.IO scheme clips the part of the residual attributed to the measurement error: $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$2 using a Huber function: $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$3 to obtain the clipped residual $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$4, where $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$5 is a threshold.

The final IO-robust estimate is: $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$6 with $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$7 (left-)invertible or replaced by a generalized inverse if necessary. This update is piece-wise linear in $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$8, with a state-dependent effective gain.

4. Choice of Clipping Threshold and Implementation

Selection of the Huber threshold $$\begin{align*} X_t &= F_t X_{t-1} + v_t, & X_t \in \mathbb{R}^p \ Y_t &= Z_t X_t + \varepsilon_t, & Y_t \in \mathbb{R}^q \end{align*}$$9 is based either on knowledge of the contamination radius $F_t$0 or on a desired loss of efficiency under ideal (non-contaminated) conditions. The minimax prescription determines $F_t$1 via: $F_t$2 or, allowing an efficiency “premium” $F_t$3,

$F_t$4

In practice, $F_t$5 is precomputed offline for the specification of $F_t$6 or $F_t$7. The run-time cost relative to the classical Kalman filter consists of a single matrix-vector product, vector norm, and a scalar minimum operation per time step.

5. Theoretical Properties and Comparative Behavior

The rLS.IO filter achieves one-step minimax-MSE optimality for IO-contamination radius $F_t$8, reaching the saddle-point in the minimax filtering risk. Under reasonable regularity conditions, recursive application ensures optimal single-step IO robustness across the time series.

Compared with the standard Kalman filter, rLS.IO rapidly tracks abrupt changes (such as level shifts or local linear trends in $F_t$9), attributing large innovations to genuine state changes rather than absorbing them into the measurement noise channel. In contrast to the AO-robust filter rLS.AO, which uses Huberization on the innovation term $Z_t$0 and thus downweights exogenous spikes in the observations, rLS.IO is explicitly designed to follow endogenous, regime-changing outliers.

In scenarios involving both AO and IO contamination, hybrid strategies (running rLS.AO and rLS.IO in parallel and switching adaptively based on run-length statistics of large innovations) can jointly capture both classes of disturbance, with performance trade-offs in response delay.

6. Context: Related Methods and Historical Perspective

Dedicated IO robustness, as formalized in rLS.IO, is compared against the ACM filter [Martin and Masreliez, 1977; Martin, 1979] and non-parametric, repeated-median-based filters [Fried et al., 2006, 2007]. The rLS.IO approach is distinctive for its explicit minimax optimality with respect to IO contamination and for its transparent analytical structure. The underlying robustness ideas are elaborated in Ruckdeschel [2001, 2010], with theoretical proofs and saddle-point results detailed in “Optimally Robust Filtering,” Fraunhofer ITWM (Ruckdeschel, 2010).

The distinction between AO-robust and IO-robust filters, as well as their interplay within hybrid schemes, underlines the nuanced dynamics of robust state estimation in the presence of both exogenous and endogenous disturbances. A plausible implication is the importance of correctly specifying the nature of outlier contamination for optimal filter design in real-world applications.