How can a global L1 exponential observer be designed for quasilinear diffusion equations?

One boundary value is the only measurement

The setting is a one-dimensional quasilinear diffusion equation on the spatial interval from 0 to 1, with Neumann boundary conditions, so the diffusive flux vanishes at both ends. The state z(x,t) is distributed over the whole interval, but the only available measurement is the state at one end of the domain, y(t) = z(1,t). The paper's motivation is that in applications such as heat transfer in high-temperature plasmas, porous-media flows, and population dynamics, the distributed state cannot be measured everywhere, so reconstruction for monitoring and control has to rely on a limited number of accessible locations such as the boundary.

The observer is a copy of the system, with the same diffusivity potential, driven only by a correction at the measured end. At the unmeasured end the observer keeps the same zero-flux condition as the plant; at the measured end the flux is set to k times the difference between the diffusivity potential evaluated at the measured boundary value and at the estimated boundary value, where k > 0 is a freely tunable observer gain. The estimation error e = z − ẑ then satisfies its own diffusion equation with a homogeneous Neumann condition at x = 0 and the boundary condition wx(1,t) = −k w(1,t), where w is the difference of diffusivity potentials.

A squared L2 Lyapunov function needs extra assumptions

The natural first attempt for a nonlinear diffusion equation is a Lyapunov analysis based on the squared L2 norm of the estimation error. The paper explains why this is not well suited to a general nonlinear diffusivity: differentiating that functional along the error dynamics produces terms involving spatial derivatives of the nonlinear diffusivity. Controlling those terms generally requires additional structural assumptions on the diffusivity, such as suitable upper bounds or growth conditions, which would restrict the generality of the result.

The paper's stated contribution is to avoid those extra assumptions. It exploits the particular structure of the nonlinear boundary correction term and constructs a Lyapunov functional from the spatial primitive of the estimation error rather than from the error itself. In the resulting dissipation identity, the nonlinear diffusivity enters only through a positive averaged derivative, so a lower bound on the diffusivity is enough to make the functional decay. This is the mechanism that keeps the result valid for a general nonlinear diffusivity rather than a restricted class.

Decay of the Lyapunov functional is not yet an L1 bound

The Lyapunov functional used in the proof is not equivalent to the L1 norm in which global stability is claimed, so its decay alone does not give the theorem. The paper addresses this gap in two steps. First, a Poincaré-type inequality with the observer gain k built in shows that the Lyapunov functional is bounded by the squared L2 norm of the error, which yields exponential decay of the functional and an integrability statement for the squared error over time.

Second, the paper derives a separate contraction property for the L1 norm of the estimation error. This is done by regularizing the absolute value function and using the fact that the nonlinear diffusivity is strictly increasing, which forces the estimation error and the difference of diffusivity potentials to have the same sign. Combining the contraction property with the decay of the Lyapunov functional produces the global exponential stability statement in L1.

The L1 bound depends on two gain-determined constants

The main theorem states that for k > 0 and under the assumption that the diffusivity is bounded below by a strictly positive constant, the estimation error satisfies an L1 inequality of the form: the integral of |e(x,t)| over the interval is at most min{1, C1 exp(−λµt)} times the integral of |eo(x)| over the interval, for all t ≥ 0. Here µ is the positive lower bound on the diffusivity, and λ and C1 are determined by the observer gain through ξ, the unique solution in (0, π/2) of ξ tan(ξ) = k, with λ = ξ² and C1 = √(exp(1)) (ξ² + ξ cot(ξ)).

The paper also characterizes the range of these constants. Because ξ lies in (0, π/2), λ lies in (0, π²/4), so the exponential rate is bounded above regardless of how large the gain is made. The paper shows that C1 lies between exp(1/2) and exp(1/2) π², and that the function ξ cot(ξ) + ξ² is strictly increasing on (0, π/2). The min{1, ·} form matters: for small times the bound reduces to the contraction property rather than to exponential decay.

Simulations confirm the rate and its saturation

The numerical example uses a diffusivity potential whose derivative is (z² + 1)^(5/4), which is at least 1, so the uniform parabolicity assumption holds. The paper notes that the resulting equation models ion temperature along a magnetic field line in the scrape-off layer of a tokamak. The plant is discretized on a fine mesh with N = 60 and time step 10⁻⁴, while the observer uses a coarser mesh with N̂ = 30 and time step 2 × 10⁻⁴, reflecting the fact that the observer is the numerically implemented system.

The first simulation varies the size of the initial estimation error through A ∈ {20, 200, 2000} at k = 1 and shows the L1 norm of the estimation error converging exponentially to zero on a semi-logarithmic scale, as the theorem predicts. The second simulation plots the ratio of the L1 error norm to the L1 norm of the initial error for k ∈ {1, 10, 20, 30} at A = 20. The convergence rate increases with k, but the improvement diminishes: the gap between k = 1 and k = 10 is larger than the gap between k = 20 and k = 30, which the paper attributes to λ being confined to (0, π²/4).

Output noise falls outside the theorem but not outside the method

A third simulation adds measurement noise to the output, with ν drawn independently at each instant from a centred Gaussian distribution of standard deviation σn ∈ {0, 10⁻², 10⁻¹}, at k = 5 and A = 2. The paper states plainly that the main theorem does not cover this case, because the theorem is proved for the noise-free output. It then notes that by adapting the proof of the L1 contraction lemma, one can show an L1 input-to-state stability result with respect to the noise, which is in agreement with the simulated behaviour.

This distinction matters for reading the result: the exponential L1 bound is a statement about the noise-free observer, while the noisy case is supported by an adapted argument and simulation rather than by the theorem as stated. The paper's conclusion also lists what remains open: global exponential stability in the Lp norm for p > 1, extension to higher-dimensional domains, and extension from classical solutions to weak solutions, the last of which the paper says would require alternative arguments at some steps of the proof.

Uniform parabolicity is the price of generality

The result holds for a general nonlinear diffusivity, but only under the assumption that the diffusivity is bounded below by a strictly positive constant, that is, that the equation is uniformly parabolic. The paper presents this as the only assumption needed on the diffusivity, in contrast to the upper bounds or growth conditions that a squared L2 Lyapunov analysis would require. The observer also requires regularity and compatibility conditions on both the plant and observer initial conditions so that classical solutions exist.

The scope is a one-dimensional spatial domain with Neumann boundary conditions and a single boundary measurement. The paper compares its contribution with the existing landscape: for constant diffusivity there is a rich body of observer design methods such as backstepping, Kalman filtering, and pole placement, while for quasilinear diffusion the only prior result the authors know of that does not linearize the equation is their own earlier work, which estimates a bounded region of attraction for a backstepping observer. The step here is from a bounded region of attraction to a global guarantee, at the cost of restricting the measurement to one boundary value and the analysis to the L1 norm.

Raising the gain helps, but with diminishing returns

The most decision-relevant feature of this paper is the relationship between the observer gain k and the guaranteed convergence rate. The rate is λµ with λ = ξ² and ξ tan(ξ) = k, so λ is trapped in (0, π²/4) no matter how large k becomes. The simulations show the practical consequence: increasing k from 1 to 10 buys more than increasing it from 20 to 30. A reader tuning this observer should therefore treat k as a parameter with a ceiling on its rate benefit rather than as a dial that can be turned up indefinitely.

The second thing to weigh is the norm. The guarantee is stated in L1, and the proof reaches it through a Lyapunov functional built from the spatial primitive of the error plus a separate contraction property, not through the squared L2 norm that would be the default choice. The paper's own next steps, Lp for p > 1 and higher-dimensional domains, mark exactly where this L1, one-dimensional guarantee stops. Within its stated scope, the result is that a single boundary measurement plus a diffusivity-potential correction is enough for global exponential estimation, provided the diffusivity stays strictly positive.

Sources used in this answer