[Advanced] — Can you steer it? Can you see it? And if both, you can put the poles anywhere you like.

Controllability

A system is controllable if you can drive the state from any initial value to any desired value in finite time, using the available inputs.

The test:

Controllable iff .

Why that matrix. is where the input pushes directly. is where that push has propagated after an instant. the instant after. The columns span everything the input can eventually reach — the reachable subspace. Full rank means it reaches everywhere. (Cayley–Hamilton is why the series stops at : higher powers add nothing new.)

What uncontrollable means physically: there’s a direction in state space you cannot influence, no matter what you do with the actuators.

  • A symmetric two-mass system driven exactly at its centre of symmetry cannot excite the antisymmetric mode
  • A satellite with a thruster pointing through the centre of mass cannot produce torque
  • A quadcopter with three rotors — four degrees of freedom to control, three inputs

If an uncontrollable mode is unstable, the system cannot be stabilised. Full stop, by any controller. That’s the strongest negative result in the subject, and it’s worth checking before designing anything.

Stabilisable is the weaker, sufficient condition: all unstable modes are controllable. Uncontrollable stable modes are usually acceptable — they decay on their own and you simply can’t influence them.

The rank test is numerically fragile. is often badly conditioned, and “rank 4” versus “rank 3 with a tiny singular value” is a judgement call. Use the smallest singular value, not the rank — it tells you how controllable, which is what you actually want to know. A system that is technically controllable but nearly not will demand enormous control effort in that direction. Prefer the PBH test for reliability: the pair is controllable iff for every eigenvalue of — and it tells you which mode is the problem.

Observability

The dual question.

A system is observable if you can determine the initial state from measuring the output over a finite interval.

Observable iff .

is what you see directly; is what you see of the state’s motion; and so on. Full rank means nothing is hidden.

Unobservable means a state can be doing something arbitrary while the output shows nothing. Measure only position and you may not distinguish two internal configurations; measure only a sum of temperatures and you can’t tell which zone is hot.

Detectable is the weaker condition: all unstable modes are observable. Which is what actually matters — an unobservable stable mode decays harmlessly; an unobservable unstable mode grows without any indication at the output until something breaks.

That’s the formal version of the pole–zero cancellation warning from note 05. Cancelling an unstable pole with a controller zero makes it unobservable. The transfer function looks perfect. The state diverges.

Duality

Every controllability result has an observability twin. It’s not a coincidence — the observer design in note 10 is literally the controller design applied to the transposed system, and it means learning one gets you the other free.

ControlEstimation
ControllabilityObservability
State feedback gain Observer gain
LQRKalman filter
StabilisableDetectable

Minimal realisation

A state-space model is minimal iff it’s both controllable and observable.

Non-minimal means the model has states the transfer function can’t see — states that are unreachable from the input, unobservable at the output, or both. Convert to a transfer function and they vanish silently.

                  observable    unobservable
 controllable   │  ← this is    │ can drive it,
                │    the TF     │ can't see it
 uncontrollable │ can see it,   │ completely
                │ can't drive   │ invisible

Only the top-left block appears in the transfer function. The other three are real dynamics that a classical model is blind to — Kalman’s decomposition, and the single clearest argument for state space.

Pole placement

Given controllability, here’s the payoff.

Full state feedback:

Substituting into :

The closed-loop dynamics are the eigenvalues of .

If is controllable, you can place the eigenvalues of anywhere you want.

That’s a remarkable result. Not “improve the response” — place the poles exactly, at whatever locations you specify. Controllability is precisely the condition for it.

Computing :

  • Ackermann’s formula — closed form, and numerically poor above about 5th order. Fine for hand calculation, fine for teaching
  • place() / robust pole assignment — what you should use. For MIMO, isn’t unique, and the extra freedom is used to make the eigenvalues as insensitive as possible to perturbation
  • By hand: match coefficients between and your desired characteristic polynomial

Choosing where to put them

The part that’s actually hard. Placement is mechanical; deciding where is the design.

Start from the specification: from the overshoot requirement, from settling time, then place a dominant pair there and the rest 5–10× further left. → Time Response

Standard patterns worth knowing:

  • Butterworth — poles evenly spaced on a semicircle of radius . Good all-round transient, no dominant mode
  • Bessel — maximally flat group delay, so minimal overshoot. Good when overshoot is the binding constraint
  • ITAE — minimises . Tabulated coefficients, good compromise response

Three warnings, and they’re where beginners come unstuck:

Fast poles cost control effort. Effort scales roughly with how far you move a pole. Demand 10× the natural bandwidth and you’ll demand torque the motor doesn’t have — and once it saturates, all your careful placement is void.

Fast poles amplify noise. High gain on a noisy measurement gives noisy actuation, which wears hardware and can excite unmodelled modes.

Pole placement says nothing about robustness. You can place poles beautifully and end up with terrible stability margins — the classical margins simply aren’t part of the calculation. This is the main practical criticism of the method, and it’s why LQR is usually preferred: same structure, but the gain comes from an optimisation with known margin guarantees, and you tune weights rather than guessing pole locations.

Integral action

State feedback alone has the same failing as proportional control: steady-state error. drives the state to zero, which isn’t the same as tracking a reference.

The fix is the same as in PID — add an integrator as an extra state:

Augment the system:

then place poles for the augmented -state system. The result is state feedback plus integral action — zero steady-state error, robust to constant disturbances and to plant gain error.

And it needs anti-windup for exactly the reason PID does. → PID Control

The catch

requires the full state.

You rarely have it. You have — a few sensors, not every state. Position but not velocity. One temperature out of five.

Two options:

Add sensors. Sometimes right, often expensive, and some states aren’t directly measurable at all.

Estimate the state from what you can measure. Which requires observability — and is the entire subject of the next note.


↑ Back to Control Theory