Kalman and Bucy position the work against the Wiener–Kolmogorov tradition without claiming an empty field. They credit nonstationary time-domain work by Shinbrot, Steeg, Pugachev and Parzen; Follin’s time-varying gain and error-variance relations; Hanson’s optimality proof; Bucy’s wider derivation; and Kalman’s combination of state transition with Hilbert-space projection. Their central conclusion converged despite different routes: instead of attacking a Wiener–Hopf integral equation directly, solve a nonlinear covariance differential equation, then obtain the filter from it.

Continuous time is not just the discrete update performed very quickly. The observed signal contains a white-noise component represented by covariance intensity and a Dirac delta. The measurement arrives as an infinitesimal information stream. The filter’s state derivative is driven by an innovation signal; the covariance changes continuously according to a Riccati flow. A discrete jump equation and a continuous differential equation can approximate one another under carefully chosen limits, but they are not interchangeable strings of symbols.

The paper promises a unified route across multivariate, stationary and nonstationary cases, finite and infinite observation intervals. It separates description from mathematics, gives rigorous derivations within its formal conventions, explores analytic Riccati solutions through Hamiltonian equations, and studies when a steady-state covariance exists and attracts other solutions. That scope is much larger than attaching Bucy’s name to a continuous-time gain formula.

Assumption A1 supplies the message model dx/dt=F(t)x+G(t)u(t), possibly time varying, driven by zero-mean white noise u with covariance intensity Q(t). The observation is z(t)=H(t)x(t)+v(t), where v is independent zero-mean white measurement noise with intensity R(t). The process and observation noises are mutually uncorrelated. A sufficiently accurate linear model is basic; the filter does not infer it merely by observing z.

Assumption A2 makes R(t) positive definite. Physically, no observation component is exact. Algebraically, R⁻¹ exists in the gain K=PH′R⁻¹. The authors say additive white noise is unnecessary for sampled processes and anticipate removing the restriction in continuous time by other means, but their presented theory uses it. Sending R to zero inside the displayed formula without treating a noiseless constraint separately leaves the theorem’s domain.

The initial process is determined in one of two ways. A3 constructs a steady-state message from excitation extending to the infinite past and needs a uniformly asymptotically stable model. A3′ instead begins measurement at a finite time t₀ with known covariance of x(t₀). The authors prefer A3′ because it avoids demanding stable message dynamics and permits cases where the unconditional message variance grows indefinitely. Theorem 1 uses A1, A2 and A3′.

The filtered estimate obeys d x̂(t|t)/dt=F(t)x̂(t|t)+K(t)ẑ(t|t), where ẑ=z−H x̂ is the innovation or observed residual. The filter first predicts what the sensor should report from its internal state, subtracts that prediction from the actual observed signal, and feeds the mismatch through K. The same structure holds when F, G, H, Q and R vary with time.

This is relation I in the paper’s five-equation summary. It says the optimal filter is itself a linear feedback dynamical system. Copying the message model supplies the open-loop evolution. Closing the innovation path corrects directions seen by H. The feedback interpretation matters: the estimate does not simply average sensor values, and the correction is not attached after the dynamics. It continuously changes the state trajectory that generates the next predicted observation.

The filter starts at zero in the paper’s centred formulation under A3′, with the initial uncertainty carried by P₀. A non-zero known mean can be handled by centring. For extrapolation to a future time t₁≥t, relation V applies the state-transition matrix Φ(t₁,t) to the current filtered estimate. The authors explicitly say no similarly simple formula is then known for interpolation to an earlier target.

Define x̃(t|t)=x(t)−x̂(t|t). Relation II gives d x̃/dt=F x̃+G u−K[v+H x̃]. The deterministic part is (F−KH)x̃. Process noise enters through G; observation noise enters with −K. Increasing gain can suppress a visible state error faster, but it also injects more measurement noise. The optimal gain is therefore a covariance-weighted compromise, not “as large as possible”.

The covariance P(t)=cov[x̃,x̃] records the second-order size and orientation of this error. Differentiate P using the error equation and white-noise covariance intensities. Terms F P+P F′ transport and stretch uncertainty. G Q G′ adds process uncertainty. The measurement correction contributes −P H′R⁻¹H P after the optimal K is substituted. The result is relation IV, the continuous matrix Riccati equation.

Because the negative information term is quadratic in P, covariance evolution is nonlinear even though the state and observation equations are linear. Two initial covariances do not generally average into the solution from their average. The gain K=P H′R⁻¹ inherits that time variation. This is the paper’s organising insight: compute P, and the remaining filter coefficients follow.

Relation IV is the computational centre: dP/dt=FP+PF′−PH′R⁻¹HP+GQG′. Given P(t₀)=P₀, Theorem 1 states that a unique non-negative covariance solution exists for all later time under A1, A2 and A3′. Relation III converts it immediately into K=PH′R⁻¹. Relation I then runs the filter; relation II characterises the error; relation V propagates an estimate to a future target through Φ.

The ordering prevents a common circular explanation. Gain is not tuned first and then used to discover covariance. The covariance implied by model, noise intensities and observations’ geometry determines the gain that minimises the declared error. Nor is a single stationary gain assumed. Time-varying parameters or an unsettled initial covariance produce a time-varying K.

The covariance equation contains n(n+1)/2 distinct scalar first-order equations because P is symmetric. That can still be demanding, but it replaces an expanding observation record or a Wiener–Hopf integral equation with fixed-dimensional state. The authors distinguish a mathematically expressed formula from a computation that can actually accept numbers; their conclusion treats practical numerical organisation as part of solving the problem.

The scalar constant-coefficient equation is P′=2aP+q−cP² with c=h²/r. It has one positive equilibrium P∞=(a+√(a²+cq))/c and one non-positive root. Covariance below P∞ rises; covariance above it falls. The exact solution follows from the ratio between distances to the two roots. The numerical trace uses fourth-order Runge–Kutta with enough steps to make the difference visible as an audit rather than hidden by the plotting routine.

This laboratory isolates covariance, the quantity that determines gain K=Ph/r. It does not simulate literal white-noise sample paths or claim that a smooth slider is a stochastic observation. The point is to expose the nonlinear balance among dynamics, process excitation and measurement information.

Theorem 1 handles finite t₀. For steady-state estimation, Theorem 3 studies the limit of the covariance solution that began from perfect knowledge P₀=0 as t₀ tends to minus infinity. The limit exists if either the message model is uniformly asymptotically stable or the model is completely observable in the stated sense. Complete observability uses an information matrix accumulated over a past interval and requires it to be positive definite.

The paper connects that matrix to Fisher information for a special estimation problem and invokes the Cramér–Rao inequality. In a constant model, observability reduces to a rank condition involving H′,F′H′,…,(F′)ⁿ⁻¹H′. The important idea is directional: every state direction must leave enough trace in observations over some interval. Finite unconditional message variance is sufficient through stability, but informative observation can support estimation even when the uncontrolled message variance does not settle.

When matrices are constant, the limiting covariance is constant and therefore an equilibrium of the Riccati equation. Theorem 5 says the stationary Wiener problem can be solved by setting the Riccati right-hand side to zero and choosing the non-negative solution, provided the conditions of Theorems 3 and 4 hold. Merely finding an algebraic root is not enough when those conditions fail; one of the paper’s examples is included to demonstrate that warning.

An optimal filter can still have unbounded response to bounded input. The authors ask for uniformly complete observability and uniformly complete controllability of the message model, positive lower and upper bounds on Q and R, and a bound on F. Under these conditions, Theorem 4 states that the optimal filter is uniformly asymptotically stable and every covariance solution starting from a symmetric non-negative P₀ converges to the limiting P̄(t).

Controllability here belongs to the process-excitation geometry through the dual model; observability belongs to the sensor geometry. Their uniform versions require information over bounded windows rather than at a single instant. The conditions are invariant under the filtering–regulation duality. This is how results about deterministic control support results about stochastic estimation without erasing the difference between their physical roles.

The convergence statement makes the variance equation a stable computational method in the authors’ theoretical sense: initial covariance errors decay under the conditions. They expect relative insensitivity to round-off and point to Lyapunov methods for convergence-rate bounds. Modern finite-precision practice still checks symmetry and non-negativity because discretisation can violate structures preserved by the exact differential equation.

The authors introduce a quadratic Hamiltonian and its canonical differential equations for paired vectors. The enlarged system is linear even though the covariance equation is nonlinear. Partition its state-transition matrix into blocks. A matrix fractional expression built from those blocks and P₀ recovers P(t), provided the denominator block is invertible. This is the continuous analogue of turning a nonlinear ratio into linear numerator-and-denominator dynamics.

The representation clarifies theory and duality, but the paper refuses to oversell it. When coefficients vary with time, computing the Hamiltonian transition matrix can be as difficult as solving the original Riccati equation. What has been established is an equivalence of analytic difficulty with a linear variable-coefficient system, not a universal closed elementary formula.

The lift also connects estimation to the calculus of variations, where Riccati equations and canonical Hamiltonian systems already occur. Under duality, covariance for filtering corresponds to the quadratic performance matrix for regulation. The shared linear lift explains part of that formal resemblance while leaving observation noise and control cost in their distinct physical places.

Section 11 uses examples to duplicate, simplify or extend earlier filtering results. Scalar cases expose stationary and nonstationary variance behaviour. Paired regulation diagrams demonstrate duality. Periodic coefficients bring Floquet theory into the covariance problem. A second-order case shows that closed expressions and limiting behaviour depend sharply on parameter conditions. The examples are analytic stress tests of the theory, not evidence that one fixed filter fits every signal.

Section 12 calls a system adaptive when it changes parameters in response to measured changes in its environment. In the ideal filter, environment changes appear as time dependence in F, G, H, Q and R. If those changes were detected instantaneously and exactly, only parameter values would change; the form of relations I–IV would remain. That supplies a benchmark against which an actual adaptive filter can be compared. It does not identify the parameters, detect regime changes or quantify delay in doing so.

The conclusion separates theoretical from computational work. Theory asks for general form, existence, physical realisability, stability and characterisation by quantities such as signal-to-noise ratio or information rate. Computation asks whether numbers can actually be substituted at moderate system order. The authors call the variance equation both practically useful and theoretically clean, while retaining open problems: non-white noise in a convenient continuous canonical form, Lyapunov study of variance equations, and deeper links to variations and information theory.

Today’s audit adds discretisation, conditioning and model criticism. Integrate P with a structure-preserving method when possible. Check innovations for temporal correlation and scale. Compare declared covariance with realised error on held-out trajectories. Report how parameter estimates were obtained. A filter may be mathematically optimal under A1–A3′ and poorly calibrated in the world because those premises fail.