Luenberger State Observer: Reconstructing Unmeasured States from Outputs

What It Is and Where It Is Used

A Luenberger observer is a dynamic system that runs alongside a plant in software (or analog hardware) and outputs an estimate x̂ of the plant's full internal state vector x. You need it because modern control laws — LQR, pole placement, model-predictive control — assume you can feed back every state, yet real hardware instruments only a handful of them. A robot arm may have encoders on the joints but not on joint velocities or motor currents; an aircraft measures pitch angle but not the aerodynamic states that produce it.

• Sensorless motor drives: field-oriented control estimates rotor flux and position from measured stator currents and voltages, eliminating a costly, failure-prone resolver.
• Robotics: reconstructs joint velocities and load torques from position encoders alone.
• Aerospace and automotive: estimates sideslip, angle of attack, or battery state-of-charge that no cheap sensor reports directly.

Because it is deterministic and computationally light — a few matrix multiplies per sample — the Luenberger observer runs comfortably inside a 10–100 kHz control loop on an ordinary microcontroller or DSP.

How It Works: The Innovation-Corrected Model

The observer is a copy of the plant model plus a correction term. The plant is ẋ = Ax + Bu, y = Cx. The observer runs:

x̂̇ = A·x̂ + B·u + L·(y − C·x̂)

Here x̂ is the estimated state, u is the same input applied to the plant, and y is the measured output. The bracketed term (y − Cx̂) is the innovation or output-prediction error — the mismatch between what the sensor reads and what the model predicts it should read. L is the n×p observer gain matrix that converts that scalar/vector mismatch into a correction spread across all n states.

Subtracting the observer from the plant gives the estimation error e = x − x̂, whose dynamics are:

ė = (A − LC)·e

The input u, the B matrix, and the plant state have all cancelled — the error evolves autonomously. If L is chosen so that (A − LC) is Hurwitz (all eigenvalues in the left half-plane), then e → 0 for any initial mismatch, and x̂ tracks the true x regardless of where the estimate started.

Key Quantities and a Worked Design

The design freedom is entirely in placing the eigenvalues of (A − LC). By duality, designing L for (A, C) is identical to designing a state-feedback gain for the dual pair (Aᵀ, Cᵀ): L = (place(Aᵀ, Cᵀ, poles))ᵀ. For single-output systems, Ackermann's formula gives L in one shot: L = φ(A)·𝒪⁻¹·[0 … 0 1]ᵀ, where 𝒪 is the observability matrix [C; CA; …; CAⁿ⁻¹] and φ(s) is the desired observer characteristic polynomial.

Example. Take a double integrator (position/velocity): A = [[0,1],[0,0]], C = [1, 0], measuring only position. Choose observer poles at s = −20 ± 20j (settling ≈ 4/20 = 0.2 s, natural frequency ≈ 28 rad/s). Matching φ(s) = s² + 40s + 800 to det(sI − (A − LC)) yields:

• L = [40, 800]ᵀ — so l₁ = 40 s⁻¹ corrects position and l₂ = 800 s⁻² reconstructs the unmeasured velocity.

Rule of thumb: place observer poles 2–6× faster than the closed-loop control poles so the estimate settles before the controller acts on it.

Designing and Operating It in Practice

A practical observer design follows a fixed recipe:

• Check observability first. Compute rank of 𝒪 = [C; CA; …; CAⁿ⁻¹]. If rank < n, some states are invisible from y and no L exists — you need detectability at minimum (unobservable modes already stable).
• Pick observer poles. Faster is more responsive but louder: doubling the observer bandwidth roughly doubles the amplification of sensor noise, because L scales with pole magnitude.
• Solve for L via pole placement, Ackermann (single-output), or the dual Riccati route.
• Discretize. Real loops run in discrete time, so implement x̂[k+1] = A_d·x̂[k] + B_d·u[k] + L_d·(y[k] − C·x̂[k]); keep the observer sample rate ≥ 10× the fastest observer pole to avoid aliasing.

The separation principle is what makes this tractable: you may design the controller gain K and the observer gain L completely independently. The combined closed-loop eigenvalues are exactly the union of the controller eigenvalues (of A − BK) and the observer eigenvalues (of A − LC) — they never interact.

Comparison to Kalman Filters and Reduced-Order Observers

The Luenberger observer's structure is identical to a steady-state Kalman filter — both are model-plus-correction estimators of the form x̂̇ = Ax̂ + Bu + L(y − Cx̂). The difference is how L is chosen. The Luenberger observer places poles by hand for a deterministic, noise-free model; the Kalman filter computes the optimal L (the Kalman gain) by solving a Riccati equation that trades process-noise covariance Q against measurement-noise covariance R. When Q and R are known, the Kalman gain is provably minimum-variance; when they are not, hand-tuned Luenberger poles are simpler and perfectly adequate.

• Reduced-order (minimal) observer: if p outputs are measured cleanly, those states need no estimation, so you build an observer of order n − p only for the hidden states — saving computation but forfeiting the noise-averaging that a full-order observer applies to measured channels.
• Sliding-mode / extended observers: handle nonlinearity or unknown inputs that the linear Luenberger form cannot.

All descend from the same 1966 Luenberger framework of identity, reduced-order, and functional observers.

Failure Modes, Trade-offs, and Significance

Observers fail in predictable ways:

• Model mismatch. The observer trusts A, B, C. If the real plant's parameters drift — motor resistance rising 40% as it heats, an unmodeled resonance — the estimate carries a steady bias the innovation cannot fully correct.
• Noise amplification. Aggressive (fast) poles make L large, so measurement noise on y is injected hard into x̂. The classic trade-off: convergence speed versus noise rejection. A Kalman filter formalizes this balance; a Luenberger design manages it by pole choice.
• Loss of observability at certain operating points (e.g., a sensorless PMSM drive near zero speed, where back-EMF vanishes) can make the error dynamics undamped and the estimate diverge.

Its historical significance is hard to overstate: by proving that state feedback and state estimation could be designed separately yet combined without penalty, Luenberger's work made full state-space control practical for systems with limited instrumentation — the enabling idea behind sensorless drives, inertial navigation, and virtually every modern embedded controller that acts on states it cannot directly see.

Luenberger observer vs. related state-estimation approaches

Property	Full-order Luenberger	Reduced-order Luenberger	Kalman filter
States estimated	All n states	n − p unmeasured states	All n states
Handles noise	No (deterministic)	No (deterministic)	Yes (optimal, stochastic)
Gain L computed from	Pole placement / Ackermann	Pole placement on reduced system	Riccati equation (Q, R)
Order of estimator	n	n − p	n
Design inputs	Desired observer poles	Desired observer poles	Noise covariances Q, R
Typical use	Deterministic LTI control	Save computation, clean sensors	Noisy sensors, sensor fusion