Bryson–Frazier Smoother

• Bryson–Frazier smoother is a two-filter algorithm that recasts the backward pass in the information domain, enhancing state estimation in linear and nonlinear models.
• It leverages a backward information filter and Riccati-like equations to compute smoothing errors without expensive covariance inversions.
• Recent advances extend its application to nonlinear, regime-switching, and factor-graph models, improving performance in sensor fusion, tracking, and navigation.

The Bryson–Frazier smoother is a two-filter smoothing algorithm for estimating the hidden state in linear (and certain nonlinear) state-space models, formulated in either discrete or continuous time. Unlike the Rauch–Tung–Striebel (RTS) smoother, which combines forward-filtering and backward-recursions using covariance-based updates, the Bryson–Frazier approach recasts the backward pass in the information domain, yielding computational, numerical, and structural advantages. Recent advances have provided rigorous derivations of the continuous-time Bryson–Frazier smoother, pathwise representations, and extensions to conditionally linear, regime-switching, and nonlinear models (Kurisaki, 5 Jan 2026, Nguyen et al., 2017, Petersen et al., 2019).

1. Foundational Framework for the Continuous-Time Bryson–Frazier Smoother

The filtering and smoothing problem is considered in the context of a continuous-time linear Gaussian model: $$\begin{aligned} dX_t &= a(t) X_t\,dt + b(t)\,dV_t, \quad X_0 \sim \mathcal N(m_0, P_0), \ dY_t &= c(t) X_t\,dt + \sigma(t)\,dW_t, \quad Y_0 = 0, \end{aligned}$$ where $X_t$ is the $d_1$-dimensional unobserved process, $Y_t$ is the $d_2$-dimensional observation, $V_t$ and $W_t$ are independent standard Brownian motions, and the time-varying system matrices $a(t), b(t), c(t), \sigma(t)$ are measurable and bounded, with $\sigma(t)\sigma(t)^\top \succ 0$ for all $t$ (Kurisaki, 5 Jan 2026).

The forward Kalman–Bucy filter computes the conditional mean and covariance of $X_t$ given $\mathcal Y_t = \sigma(Y_s\, :\, 0 \le s \le t)$, while the smoothing distribution $p(x_s \mid \mathcal Y_t)$ seeks to condition on all observations up to terminal time $t$ for each $0 \le s \le t$.

2. Pathwise Representation and Smoothing Error Dynamics

A key insight is the pathwise representation of the smoothing error: for fixed $t$, define a backward Riccati-like matrix $\phi(s;t)$ as the unique negative-semidefinite solution of

$$\frac{\partial}{\partial s} \phi(s;t) = -\phi\,b\,b^T\,\phi - a^T\,\phi - \phi\,a + c^T (\sigma\sigma^T)^{-1} c, \quad \phi(t;t) = 0.$$

An auxiliary process $\xi_{s;t}$ is defined on $0 \le s \le t$ by

$$d_s\,\xi_{s;t} = (a(s) + b(s)b(s)^T \phi(s;t)) \xi_{s;t}\,ds + b(s)\,d\widetilde V_s,$$

where $\widetilde V$ is independent Brownian motion, and the initial covariance $V[\xi_{0;t}]$ is specified so that $$\operatorname{Cov}(\xi_{s;t}, \xi_{u;t}) = \gamma(s,u;t)$$ matches the smoothing error covariance.

The main theorem states that $X_s - E[X_s|\mathcal Y_t]$ has the same law as $\xi_{s;t}$, i.e., the smoothing error is Ornstein–Uhlenbeck–type. The full smoothing mean and covariance are given by

$$\boxed{ \mu_{s;t} = E[X_s] + \int_0^t \gamma(s,u;t)\, c(u)^T (\sigma(u)\sigma(u)^T)^{-1} (dY_u - c(u)E[X_u]\,du) }$$

$$\boxed{ \operatorname{Cov}(X_s, X_u \mid \mathcal Y_t) = \gamma(s,u;t) }$$

(Kurisaki, 5 Jan 2026).

3. Bryson–Frazier Backward Recursion

The pathwise mean formula is exact but not a recursive equation. The Bryson–Frazier smoother introduces a backward “information” process $\rho_{s;t}$: $$\rho_{s;t} = \int_s^t \alpha(s,u;t)^T c(u)^T (\sigma(u)\sigma(u)^T)^{-1} (dY_u - c(u)E[X_u]\,du),$$ where $\alpha(s,u;t)$ solves

$$\frac{\partial}{\partial u}\alpha(s,u;t) = (a(u) + b(u)b(u)^T \phi(u;t))\alpha(s,u;t), \quad \alpha(s,s;t) = I_{d_1}.$$

The smoothed mean then solves the backward ODE–SDE: $$\begin{aligned} d_s\,\mu_{s;t} &= \left\{a(s) + b(s)b(s)^T \phi(s;t)\right\}\mu_{s;t}\,ds - b(s)b(s)^T \phi(s;t) E[X_s]\,ds \ &\quad + b(s)b(s)^T \rho_{s;t}\,ds, \ \mu_{t;t} &= E[X_t] + \gamma(t,t;t) c(t)^T(\sigma(t)\sigma(t)^T)^{-1}(dY_t - c(t)E[X_t]dt). \end{aligned}$$ The backward Riccati $\phi$, mean $\mu$, and $\rho$ are integrated backward in $s$ from $t$ to $0$ (Kurisaki, 5 Jan 2026).

This yields the Bryson–Frazier cooler for the continuous-time linear model, which contains as particular cases the Kalman–Bucy (filtering) and RTS (smoothing) equations.

4. Two-Filter Smoothers and Information Formulation

In discrete time, the Bryson–Frazier smoother is best characterized as a two-filter approach. One runs:

• A forward Kalman filter producing $p(x_t|y_{1:t}) = \mathcal N(\hat x_{t|t}, P_{t|t})$,
• A backward information filter with information parameters $(\Lambda_t, \eta_t)$:

$$\begin{aligned} \Lambda_t &= H_{t+1}^T R_{t+1}^{-1} H_{t+1} + F_{t+1}^T \Lambda_{t+1} F_{t+1}, \ \eta_t &= H_{t+1}^T R_{t+1}^{-1} y_{t+1} + F_{t+1}^T \eta_{t+1}. \end{aligned}$$

The smoothed distribution is computed by combining the forward Gaussian parameters with the backward information parameters using closed-form expressions (Nguyen et al., 2017).

This structure naturally generalizes to hybrid models with conditionally linear–Gaussian structure, as well as to Rao–Blackwellized SMC smoothers using the two-filter Bryson–Frazier formalism.

5. Extensions to Nonlinear, Regime-Switching, and Factor-Graph Settings

For nonlinear state-space models, the Bryson–Frazier methodology underpins the modified Bryson–Frazier (MBF) smoother, which extends to the setting of approximate Gaussian message passing on Forney-style factor graphs. The MBF smoother computes forward and backward Gaussian messages through deterministic nonlinear nodes using numerical quadrature (e.g., Unscented Transform, cubature, Gauss–Hermite) for the forward pass and a Rauch–Tung–Striebel-type backward pass: $$\begin{aligned} \tilde\Lambda_{i-1} &= \Lambda_{i-1} C_{i-1,i}\tilde\Lambda_i C_{i-1,i}^T \Lambda_{i-1},\ \tilde\xi_{i-1} &= \Lambda_{i-1} C_{i-1,i} \tilde\xi_i,\ \Lambda_{i-1}^{\mathrm{smooth}} &= \Lambda_{i-1} + \tilde\Lambda_{i-1},\ \xi_{i-1}^{\mathrm{smooth}} &= \xi_{i-1} + \tilde\xi_{i-1}. \end{aligned}$$ Only one matrix inversion is required per step in the backward pass, yielding computational efficiency and numerical robustness (Petersen et al., 2019).

The Bryson–Frazier and MBF smoothers are suitable for large-scale sensor fusion, SLAM, navigation, and tracking contexts where regime switching, conditional linearity, or model nonlinearity are encountered.

6. Pathwise Sampling and Monte Carlo Evaluation

The pathwise representation enables exact sampling from the smoothing distribution. The recipe is:

1. Solve the forward Riccati and filter mean;
2. Integrate the backward Riccati $\phi$ to obtain the necessary backward information;
3. Simulate an auxiliary Brownian motion and initial Gaussian variable for the Ornstein–Uhlenbeck-type error process $\xi_{s;t}$;
4. Evolve $\xi_{s;t}$ via its SDE;
5. Form smoothed sample paths $X^{(i)}_s = \mu_{s;t} + \xi_{s;t}$.

Each trajectory is distributed exactly according to the smoothing law, enabling estimation of path-dependent functionals, Monte Carlo EM Q-functions, and construction of confidence bands (Kurisaki, 5 Jan 2026).

7. Computational Properties and Implementation Considerations

The Bryson–Frazier smoother avoids explicit inversion of state covariances and requires the integration of a small set of ordinary and stochastic differential equations (three matrix-valued ODEs and one linear SDE in continuous time). Standard solvers are applicable. The only invertibility assumption pertains to the observation noise covariance $\sigma(t)\sigma(t)^T \succ 0$, which ensures equivalence of measures and positiveness of $P(t)$ (Kurisaki, 5 Jan 2026).

In discrete-time regime-switching models, computational complexity is $O(N J)$ per time step for Rao–Blackwellized SMC smoothers with $N$ particles and $J$ regimes. The use of particle rejuvenation in the backward pass, in tandem with the Bryson–Frazier formalism, improves variance and accuracy in smoothing with complex regime-switching models (Nguyen et al., 2017).

---

References:

• (Kurisaki, 5 Jan 2026) "Pathwise Representation of the Smoothing Distribution in Continuous-Time Linear Gaussian Models"
• (Nguyen et al., 2017) "Particle rejuvenation of Rao-Blackwellized Sequential Monte Carlo smoothers for Conditionally Linear and Gaussian models"
• (Petersen et al., 2019) "On Approximate Nonlinear Gaussian Message Passing On Factor Graphs"