CHAPTER 34 / STATE SPACE

Solving state equations: eigenvalues, transfer functions, and the matrix exponential

Given A, B, C and D, you can find the poles, the transfer function and the complete time response by hand. This chapter works through each calculation.

THE SHORT ANSWER

Solving a linear state equation ẋ = Ax + Bu means finding x(t) = e^{At}x(0) plus a convolution of past inputs, where the state-transition matrix e^{At} carries the state forward in time. The same model gives the transfer function G(s) = C(sI − A)⁻¹B + D, whose poles are eigenvalues of A unless a numerator factor cancels them. e^{At} is found by inverse Laplace transform, diagonalisation or the Cayley–Hamilton method.

x(t)=eAtx(0)+∫0teA(t−τ)Bu(τ) dτ,G(s)=C(sI−A)−1B+Dx(t)=e^{At}x(0)+\int_0^t e^{A(t-\tau)}Bu(\tau)\,d\tau,\qquad G(s)=C(sI-A)^{-1}B+D
WHAT YOU WILL LEARN

Compute the eigenvalues, the transfer function, e^{At} and the full time response of a second-order state model by hand, and write a state model for a given transfer function.

Early in the course we wrote a system as ẋ = Ax + Bu, y = Cx + Du, and then spent the classical chapters working with transfer functions. We now reopen that state model and compute with it. The same four matrices give the poles, the transfer function and the time response, including the part caused by the initial state.

Review the previous lesson: Thermostats, PI and heating limits
START WITH AN EXAMPLE

First, picture it

Start with one state. The equation ẋ = ax + bu has the solution x(t) = e^{at}x(0) plus a weighted sum of past inputs, and the sign of a decides whether the free motion dies out or grows. A state equation with n states has the same solution with the matrix A in place of the number a. The matrix exponential e^{At} replaces e^{at}, and the eigenvalues of A take over the role of a: each eigenvalue λ contributes a term e^{λt} to the motion. Pull a mass on a spring away from rest and let go. The motion you see is e^{At} acting on the initial position and velocity.

THE IDEA

Build the idea step by step

Take the Laplace transform of ẋ = Ax + Bu and keep the initial state: sX(s) − x(0) = AX(s) + BU(s). Multiply by (sI − A)⁻¹ to solve for X(s). The result has two parts, one driven by x(0) and one driven by the input. Set x(0) = 0 and substitute into y = Cx + Du to get the transfer function G(s) = C(sI − A)⁻¹B + D. Every entry of (sI − A)⁻¹ is a polynomial divided by det(sI − A), so the poles of G(s) are roots of det(sI − A) = 0, and those roots are the eigenvalues of A. A pole is missing from G(s) only when a numerator factor cancels it. Transform (sI − A)⁻¹ back to time and you have the state-transition matrix Φ(t) = e^{At}. It carries any initial state forward to time t and, inside a convolution integral, gives the response to any input.

Read the equations

x˙(t)=Ax(t)+Bu(t),y(t)=Cx(t)+Du(t)\dot x(t)=Ax(t)+Bu(t),\qquad y(t)=Cx(t)+Du(t)

The state equation gives the rate of change of every state; the output equation says what is measured. For n states, one input and one output, A is n×n, B is n×1, C is 1×n and D is a number.

X(s)=(sI−A)−1x(0)+(sI−A)−1BU(s)X(s)=(sI-A)^{-1}x(0)+(sI-A)^{-1}BU(s)

This is the Laplace transform of the state equation, solved for X(s). The first term is the zero-input response caused by the initial state, and the second is the zero-state response caused by the input.

G(s)=Y(s)U(s)∣x(0)=0=C(sI−A)−1B+DG(s)=\frac{Y(s)}{U(s)}\bigg|_{x(0)=0}=C(sI-A)^{-1}B+D

Set the initial state to zero and substitute X(s) into the output equation. This one formula converts any state model into its transfer function.

(sI−A)−1=adj⁡(sI−A)det⁡(sI−A),det⁡(sI−A)=0 ⟺ s=λi(sI-A)^{-1}=\frac{\operatorname{adj}(sI-A)}{\det(sI-A)},\qquad \det(sI-A)=0\ \Longleftrightarrow\ s=\lambda_i

The determinant is the common denominator of every entry, so the eigenvalues of A are the candidates for the poles of G(s). They are exactly the poles when no numerator factor cancels one of them.

[abcd]−1=1ad−bc[d−b−ca]\begin{bmatrix}a&b\\c&d\end{bmatrix}^{-1}=\frac{1}{ad-bc}\begin{bmatrix}d&-b\\-c&a\end{bmatrix}

The 2×2 inverse you need in every hand calculation: swap the diagonal, negate the off-diagonal, divide by the determinant.

G(s)=b2s2+b1s+b0s3+a2s2+a1s+a0A=[010001−a0−a1−a2],B=[001],C=[b0b1b2],D=0\begin{aligned}G(s)&=\frac{b_2s^2+b_1s+b_0}{s^3+a_2s^2+a_1s+a_0}\\[4pt] A&=\begin{bmatrix}0&1&0\\0&0&1\\-a_0&-a_1&-a_2\end{bmatrix},\quad B=\begin{bmatrix}0\\0\\1\end{bmatrix},\quad C=\begin{bmatrix}b_0&b_1&b_2\end{bmatrix},\quad D=0\end{aligned}

Controllable canonical form for a third-order transfer function. The denominator coefficients fill the bottom row of A with reversed order and opposite sign, and the numerator coefficients fill C. The same pattern extends to any order.

Ao=AcT,Bo=CcT,Co=BcTA_o=A_c^{\mathsf T},\qquad B_o=C_c^{\mathsf T},\qquad C_o=B_c^{\mathsf T}

Observable canonical form is the dual of the controllable form (subscript c): transpose A and exchange B and C. Both forms have the same transfer function.

Φ(t)=eAt=I+At+A2t22!+⋯=L−1 ⁣{(sI−A)−1}\Phi(t)=e^{At}=I+At+\frac{A^2t^2}{2!}+\cdots=\mathcal{L}^{-1}\!\left\{(sI-A)^{-1}\right\}

The state-transition matrix. The series defines it, and the inverse Laplace transform is how you compute it by hand. With no input, x(t) = Φ(t)x(0).

Φ(0)=I,Φ(t1+t2)=Φ(t1)Φ(t2),Φ−1(t)=Φ(−t),Φ˙(t)=AΦ(t)\Phi(0)=I,\qquad \Phi(t_1+t_2)=\Phi(t_1)\Phi(t_2),\qquad \Phi^{-1}(t)=\Phi(-t),\qquad \dot\Phi(t)=A\Phi(t)

Properties of the state-transition matrix. Moving the state by t₁ and then by t₂ is the same as moving it by t₁ + t₂, and running time backwards undoes the motion. Use Φ(0) = I to check every e^{At} you compute.

x(t)=eAtx(0)+∫0teA(t−τ)Bu(τ) dτx(t)=e^{At}x(0)+\int_0^t e^{A(t-\tau)}Bu(\tau)\,d\tau

The complete solution of the state equation. The input at each past time τ kicks the state by Bu(τ), and e^{A(t−τ)} carries that kick forward to time t. The output is y(t) = Cx(t) + Du(t).

A=PΛP−1 ⟹ eAt=P[eλ1t00eλ2t]P−1,P=[v1v2]A=P\Lambda P^{-1}\ \Longrightarrow\ e^{At}=P\begin{bmatrix}e^{\lambda_1t}&0\\0&e^{\lambda_2t}\end{bmatrix}P^{-1},\qquad P=\begin{bmatrix}v_1&v_2\end{bmatrix}

Diagonalisation. The columns of P are eigenvectors, Av_i = λ_i v_i. It needs as many independent eigenvectors as states, which is guaranteed when the eigenvalues are distinct.

eAt=α0(t)I+α1(t)A,eλit=α0(t)+α1(t)λi(i=1,2)e^{At}=\alpha_0(t)I+\alpha_1(t)A,\qquad e^{\lambda_it}=\alpha_0(t)+\alpha_1(t)\lambda_i\quad(i=1,2)

Cayley–Hamilton method for two states. Solve the two scalar equations for α₀ and α₁, then build the matrix. With n states the sum runs up to Aⁿ⁻¹.

Re⁡λi(A)<0  for every i⟺eAt→0  as t→∞\operatorname{Re}\lambda_i(A)<0\ \text{ for every } i\quad\Longleftrightarrow\quad e^{At}\to 0\ \text{ as } t\to\infty

Stability test. Every entry of e^{At} is a sum of terms e^{λt}, possibly multiplied by powers of t, so the free response dies out exactly when all eigenvalues lie in the open left half-plane.

x=Txˉ:Aˉ=T−1AT,Bˉ=T−1B,Cˉ=CT,Cˉ(sI−Aˉ)−1Bˉ=C(sI−A)−1Bx=T\bar x:\qquad \bar A=T^{-1}AT,\quad \bar B=T^{-1}B,\quad \bar C=CT,\qquad \bar C(sI-\bar A)^{-1}\bar B=C(sI-A)^{-1}B

Similarity transformation. A change of state coordinates changes the matrices but keeps the eigenvalues and the transfer function.

What the symbols mean

x(t)x(t)
state vector with n entries; each entry has its own unit
u(t), y(t)u(t),\ y(t)
input and output signals, in their physical units
A,B,C,DA,B,C,D
system, input, output and direct-feedthrough matrices; constant
x(0)x(0)
initial state at t = 0
t, τt,\ \tau
time and the integration variable for past time, s
ss
Laplace variable, 1/s
II
identity matrix of the same size as A
G(s)G(s)
transfer function, output unit per input unit
Φ(t)=eAt\Phi(t)=e^{At}
state-transition matrix (matrix exponential), n×n
λi\lambda_i
eigenvalues of A, 1/s; the imaginary part is in rad/s
vi, Pv_i,\ P
eigenvectors of A and the matrix that has them as columns
Λ\Lambda
diagonal matrix of the eigenvalues
ai, bia_i,\ b_i
denominator and numerator coefficients of G(s)
α0(t), α1(t)\alpha_0(t),\ \alpha_1(t)
time functions that multiply I and A in the Cayley–Hamilton method
T, xˉT,\ \bar x
invertible coordinate-change matrix and the new state vector
nn
number of states

These formulas hold for a linear time-invariant model, where A, B, C and D are constant. The transfer function describes the zero-state response only; the solution formula also includes the initial state. All time functions here are defined for t ≥ 0.

WORKED EXAMPLE

One second-order system from start to finish

  1. Take this model. The output is the first state, and the input enters the second state equation.

    A=[01−2−3],B=[01],C=[10],D=0A=\begin{bmatrix}0&1\\-2&-3\end{bmatrix},\qquad B=\begin{bmatrix}0\\1\end{bmatrix},\qquad C=\begin{bmatrix}1&0\end{bmatrix},\qquad D=0
  2. Find the eigenvalues from the characteristic equation. Both are negative, so the system is stable. Check: their sum −3 is the trace of A and their product 2 is det A.

    det⁡(sI−A)=∣s−12s+3∣=s2+3s+2=(s+1)(s+2) ⟹ λ1=−1, λ2=−2\det(sI-A)=\begin{vmatrix}s&-1\\2&s+3\end{vmatrix}=s^2+3s+2=(s+1)(s+2)\ \Longrightarrow\ \lambda_1=-1,\ \lambda_2=-2
  3. Invert sI − A with the 2×2 rule.

    (sI−A)−1=1(s+1)(s+2)[s+31−2s](sI-A)^{-1}=\frac{1}{(s+1)(s+2)}\begin{bmatrix}s+3&1\\-2&s\end{bmatrix}
  4. Method 1, inverse Laplace transform. Expand each entry in partial fractions.

    s+3(s+1)(s+2)=2s+1−1s+2,1(s+1)(s+2)=1s+1−1s+2,−2(s+1)(s+2)=−2s+1+2s+2,s(s+1)(s+2)=−1s+1+2s+2\begin{aligned}\frac{s+3}{(s+1)(s+2)}&=\frac{2}{s+1}-\frac{1}{s+2},&\qquad \frac{1}{(s+1)(s+2)}&=\frac{1}{s+1}-\frac{1}{s+2},\\[4pt] \frac{-2}{(s+1)(s+2)}&=\frac{-2}{s+1}+\frac{2}{s+2},&\qquad \frac{s}{(s+1)(s+2)}&=\frac{-1}{s+1}+\frac{2}{s+2}\end{aligned}
  5. Transform term by term to get the state-transition matrix. At t = 0 it equals I, as it must. At t = 1 s it is [[0.6004, 0.2325], [−0.4651, −0.0972]].

    eAt=[2e−t−e−2te−t−e−2t−2e−t+2e−2t−e−t+2e−2t]e^{At}=\begin{bmatrix}2e^{-t}-e^{-2t}&e^{-t}-e^{-2t}\\-2e^{-t}+2e^{-2t}&-e^{-t}+2e^{-2t}\end{bmatrix}
  6. Method 2, diagonalisation. The eigenvectors of this companion matrix are [1, λ]ᵀ. Multiplying out the three matrices reproduces the result of method 1.

    P=[11−1−2],P−1=[21−1−1],eAt=P[e−t00e−2t]P−1P=\begin{bmatrix}1&1\\-1&-2\end{bmatrix},\quad P^{-1}=\begin{bmatrix}2&1\\-1&-1\end{bmatrix},\quad e^{At}=P\begin{bmatrix}e^{-t}&0\\0&e^{-2t}\end{bmatrix}P^{-1}
  7. Method 3, Cayley–Hamilton. Write one scalar equation per eigenvalue, solve for α₀ and α₁, and assemble α₀I + α₁A. The result is again the same matrix.

    e−t=α0−α1,e−2t=α0−2α1 ⟹ α1=e−t−e−2t,α0=2e−t−e−2teAt=α0I+α1A=[α0α1−2α1α0−3α1]\begin{aligned}e^{-t}&=\alpha_0-\alpha_1,\quad e^{-2t}=\alpha_0-2\alpha_1\ \Longrightarrow\ \alpha_1=e^{-t}-e^{-2t},\quad \alpha_0=2e^{-t}-e^{-2t}\\[4pt] e^{At}&=\alpha_0I+\alpha_1A=\begin{bmatrix}\alpha_0&\alpha_1\\-2\alpha_1&\alpha_0-3\alpha_1\end{bmatrix}\end{aligned}
  8. Transfer function. With B = [0, 1]ᵀ and C = [1, 0], the product C(sI − A)⁻¹B picks the top-right entry of the inverse. The poles −1 and −2 are the eigenvalues.

    G(s)=C(sI−A)−1B=1(s+1)(s+2)=1s2+3s+2G(s)=C(sI-A)^{-1}B=\frac{1}{(s+1)(s+2)}=\frac{1}{s^2+3s+2}
  9. Unit-step response from zero initial state. Put u = 1 into the convolution integral; Ce^{A(t−τ)}B is the top-right entry of e^{A(t−τ)}. The response reaches 0.1998 at t = 1 s and 0.3738 at t = 2 s, and settles at G(0) = 1/2.

    y(t)=∫0t(e−(t−τ)−e−2(t−τ))dτ=12−e−t+12e−2ty(t)=\int_0^t\left(e^{-(t-\tau)}-e^{-2(t-\tau)}\right)d\tau=\frac12-e^{-t}+\frac12e^{-2t}
  10. Add an initial state x(0) = [1, 0]ᵀ. The zero-input output is the top-left entry of e^{At}, and by superposition it adds to the step response. At t = 1 s the total is 0.6004 + 0.1998 = 0.8002; the final value is still 1/2.

    y(t)=2e−t−e−2t⏟zero input+12−e−t+12e−2t⏟zero state=12+e−t−12e−2ty(t)=\underbrace{2e^{-t}-e^{-2t}}_{\text{zero input}}+\underbrace{\tfrac12-e^{-t}+\tfrac12e^{-2t}}_{\text{zero state}}=\frac12+e^{-t}-\frac12e^{-2t}
More detail and model assumptions
A CLOSER LOOK

Three routes to e^{At}, and when to use each

Inverse Laplace transform always works. Form sI − A, invert it, expand each entry in partial fractions and transform term by term. For a 2×2 matrix the inverse is quick: swap the diagonal entries, change the sign of the off-diagonal entries and divide by the determinant. A repeated eigenvalue λ produces a term 1/(s − λ)², which transforms to t·e^{λt}.

Diagonalisation is fastest when the eigenvalues are distinct. Put the eigenvectors in the columns of P; then A = PΛP⁻¹ and e^{At} = P·diag(e^{λ₁t}, …, e^{λₙt})·P⁻¹. When A is in the companion form used in this chapter, the eigenvector for λ is [1, λ]ᵀ for two states and [1, λ, λ²]ᵀ for three, so you can write P without solving anything. The method fails when a repeated eigenvalue has too few independent eigenvectors, as in A = [[0, 1], [−4, −4]].

The Cayley–Hamilton theorem says that A satisfies its own characteristic equation. Every power Aᵏ with k ≥ n is therefore a combination of I, A, …, Aⁿ⁻¹, and the infinite series for e^{At} collapses to α₀(t)I + α₁(t)A + … + αₙ₋₁(t)Aⁿ⁻¹. Each eigenvalue gives one scalar equation e^{λt} = α₀ + α₁λ + … for the unknown coefficients. For a repeated eigenvalue, differentiate that equation with respect to λ to get the extra equation you need.

A CLOSER LOOK

From a transfer function to state equations

Write G(s) = N(s)/D(s) with a monic denominator, and split it into two steps: W(s) = U(s)/D(s), then Y(s) = N(s)W(s). Choose the states x₁ = w, x₂ = ẇ, x₃ = ẅ. The first rows of the state equation only say that each state is the derivative of the one before it. The last row is the differential equation for w solved for its highest derivative, which puts −a₀, −a₁, −a₂ in the bottom row of A and 1 at the bottom of B. The numerator coefficients become C = [b₀ b₁ b₂]. This is the controllable canonical form.

The observable canonical form is the dual: transpose A and exchange the roles of B and C. It has the same transfer function because G(s) is a scalar and a scalar equals its own transpose: C(sI − A)⁻¹B = Bᵀ(sI − Aᵀ)⁻¹Cᵀ. Both forms need the numerator degree to be lower than the denominator degree. If the degrees are equal, divide first and put the constant quotient in D.

Textbooks number the states in different orders, so the coefficient row may appear at the top instead of the bottom. Every such version is correct. To check any realisation, compute C(sI − A)⁻¹B + D, or at least compare det(sI − A) with D(s) and the DC gain −CA⁻¹B + D with G(0).

A CLOSER LOOK

State coordinates are not unique

Pick any invertible matrix T and define a new state by x = T·x̄. The new model has Ā = T⁻¹AT, B̄ = T⁻¹B, C̄ = CT and the same D. This change of coordinates is a similarity transformation. It keeps the eigenvalues, because det(sI − T⁻¹AT) = det(T⁻¹(sI − A)T) = det(sI − A), and it keeps the transfer function, because the factors T and T⁻¹ cancel in C̄(sI − Ā)⁻¹B̄. What changes is the meaning of each state. Position and velocity, the controllable canonical states and the modal states that make A diagonal all describe one and the same system.

Two consequences matter in an exam. First, a transfer function has no single correct state model, so an answer in a different canonical form can be fully right. Second, diagonalisation is this same transformation with T = P: in modal coordinates each state obeys its own scalar equation, which is why e^{Λt} is diagonal.

A CLOSER LOOK

Quick checks that catch arithmetic slips

Check the eigenvalues before going further: their sum equals the trace of A and their product equals det A. Check e^{At} at t = 0, where it must equal I, and check its derivative at t = 0, which must equal A. Both checks take seconds and catch most sign errors in the partial fractions.

For a stable system, the unit-step response must settle at the DC gain G(0) = −CA⁻¹B + D. The initial state adds terms that decay; it does not change that final value. If your y(t) tends to a different constant, the error is in the zero-state part.

TRY IT

Connect the example to the lab

The state-feedback lab uses a mass–spring–damper with x = [position, velocity], A = [[0, 1], [−4, −0.8]], B = [0, 1]ᵀ and C = [1, 0]. Its open-loop eigenvalues come from s² + 0.8s + 4 = 0, which gives s = −0.4 ± j1.96 s⁻¹. State feedback u = −Kx + Nr replaces A by A − BK, whose bottom row is [−4 − k₁, −0.8 − k₂]. With the default ωn = 2 rad/s and ζ = 0.7 the lab shows K = [0, 2], so det(sI − (A − BK)) = s² + 2.8s + 4 and the eigenvalues are −1.4 ± j1.43 s⁻¹. Compute these by hand first, then compare them with the controller poles and polynomial printed in the lab's model panel.

Check the eigenvalues of A − BK in the lab
YOUR EXPERIMENT

Make a prediction before moving a control.

  1. PREDICTBefore you open the lab, compute the eigenvalues of A − BK for ωn = 4 rad/s and ζ = 0.7, where K = [12, 4.8]. How much faster should the oscillation envelope decay than at ωn = 2 rad/s?
  2. CHANGE ONE THINGPin the default run with ωn = 2 rad/s, then change only ωn to 4 rad/s. Keep ζ = 0.7, the 0.2 m target and true-state feedback.
  3. OBSERVERead K and the controller poles in the model panel: −2.8 ± j2.86 s⁻¹ against −1.4 ± j1.43 s⁻¹. Check that the position settles in about half the time and that the first requested force rises from 0.8 N to 3.2 N.
  4. EXPLAINThe real part of the eigenvalues sets the envelope e^{−ζωn·t}: going from −1.4 to −2.8 s⁻¹ halves its time constant from 0.71 s to 0.36 s. The first force is N·r = ωn²·0.2, so the faster eigenvalues cost four times the force. If the force clips at its limit, the loop is no longer the linear model and the eigenvalues of A − BK stop describing it.
BE CAREFUL

Do not judge stability from G(s) alone. When a pole–zero cancellation removes an eigenvalue from the transfer function, that mode still moves inside the system, so check the eigenvalues of A. Also, e^{At} is not the matrix whose entries are e^{a_ij t}; exponentiating entry by entry gives a wrong answer for any A that is not diagonal.

Check your understanding+

A has eigenvalues −1 and +2, yet the transfer function simplifies to G(s) = 1/(s + 1). Is the system stable?

No. The eigenvalue +2 was cancelled in G(s), but the mode e^{2t} is still part of e^{At} and grows inside the state. Internal stability needs every eigenvalue of A in the open left half-plane.
PRACTICE PROBLEMS

Solve these before opening the solutions

Work each problem on paper first. Open the answer to check yours, then the full solution if you need the steps, and mark how it went.

  1. Find the eigenvalues of A and decide whether ẋ = Ax is stable.

    A=[1230]A=\begin{bmatrix}1&2\\3&0\end{bmatrix}
    Show answer

    AnswerThe eigenvalues are 3 and −2. The system is unstable because one eigenvalue is positive.

    λ1=3,λ2=−2\lambda_1=3,\qquad \lambda_2=-2
    Show full solution

    Solution

    1. Form the characteristic polynomial.

      det⁡(sI−A)=∣s−1−2−3s∣=s(s−1)−6=s2−s−6\det(sI-A)=\begin{vmatrix}s-1&-2\\-3&s\end{vmatrix}=s(s-1)-6=s^2-s-6
    2. Factor it. Check: the sum 1 is the trace of A and the product −6 is det A.

      s2−s−6=(s−3)(s+2)s^2-s-6=(s-3)(s+2)
    3. The eigenvalue 3 gives a mode that grows without bound, so the system is unstable even though the other mode decays.

      eAt contains e3t and e−2te^{At}\ \text{contains}\ e^{3t}\ \text{and}\ e^{-2t}
  2. Find the transfer function G(s) of this model and its DC gain.

    A=[−110−3],B=[02],C=[10],D=0A=\begin{bmatrix}-1&1\\0&-3\end{bmatrix},\qquad B=\begin{bmatrix}0\\2\end{bmatrix},\qquad C=\begin{bmatrix}1&0\end{bmatrix},\qquad D=0
    Show answer

    AnswerG(s) = 2/((s + 1)(s + 3)); the DC gain is 2/3.

    G(s)=2s2+4s+3,G(0)=23G(s)=\frac{2}{s^2+4s+3},\qquad G(0)=\frac23
    Show full solution

    Solution

    1. A is triangular, so its eigenvalues are the diagonal entries −1 and −3.

      sI−A=[s+1−10s+3],det⁡(sI−A)=(s+1)(s+3)sI-A=\begin{bmatrix}s+1&-1\\0&s+3\end{bmatrix},\qquad \det(sI-A)=(s+1)(s+3)
    2. Invert with the 2×2 rule.

      (sI−A)−1=1(s+1)(s+3)[s+310s+1](sI-A)^{-1}=\frac{1}{(s+1)(s+3)}\begin{bmatrix}s+3&1\\0&s+1\end{bmatrix}
    3. Multiply by B on the right, then by C on the left. C keeps the first row only.

      (sI−A)−1B=1(s+1)(s+3)[22(s+1)] ⟹ G(s)=2(s+1)(s+3)(sI-A)^{-1}B=\frac{1}{(s+1)(s+3)}\begin{bmatrix}2\\2(s+1)\end{bmatrix}\ \Longrightarrow\ G(s)=\frac{2}{(s+1)(s+3)}
    4. Set s = 0 for the DC gain. A unit step settles at 2/3.

      G(0)=21⋅3=23G(0)=\frac{2}{1\cdot 3}=\frac23
  3. Write the controllable canonical form of G(s), then the observable canonical form. Find the poles and the DC gain.

    G(s)=2s+5s3+4s2+6s+3G(s)=\frac{2s+5}{s^3+4s^2+6s+3}
    Show answer

    AnswerThe bottom row of A is [−3, −6, −4] and C = [5, 2, 0]. The poles are −1 and −1.5 ± j0.866, and the DC gain is 5/3.

    A=[010001−3−6−4],B=[001],C=[520]A=\begin{bmatrix}0&1&0\\0&0&1\\-3&-6&-4\end{bmatrix},\quad B=\begin{bmatrix}0\\0\\1\end{bmatrix},\quad C=\begin{bmatrix}5&2&0\end{bmatrix}
    Show full solution

    Solution

    1. Read the coefficients. The s² term is missing from the numerator, so b₂ = 0.

      a0=3, a1=6, a2=4;b0=5, b1=2, b2=0a_0=3,\ a_1=6,\ a_2=4;\qquad b_0=5,\ b_1=2,\ b_2=0
    2. Fill the controllable canonical pattern: −a₀, −a₁, −a₂ in the bottom row of A and b₀, b₁, b₂ in C.

      Ac=[010001−3−6−4],Bc=[001],Cc=[520]A_c=\begin{bmatrix}0&1&0\\0&0&1\\-3&-6&-4\end{bmatrix},\quad B_c=\begin{bmatrix}0\\0\\1\end{bmatrix},\quad C_c=\begin{bmatrix}5&2&0\end{bmatrix}
    3. For the observable canonical form, transpose A and exchange B and C.

      Ao=[00−310−601−4],Bo=[520],Co=[001]A_o=\begin{bmatrix}0&0&-3\\1&0&-6\\0&1&-4\end{bmatrix},\quad B_o=\begin{bmatrix}5\\2\\0\end{bmatrix},\quad C_o=\begin{bmatrix}0&0&1\end{bmatrix}
    4. The poles are the roots of the denominator. Try s = −1: −1 + 4 − 6 + 3 = 0, so divide out (s + 1). All three real parts are negative, so the system is stable.

      s3+4s2+6s+3=(s+1)(s2+3s+3) ⟹ s=−1,  s=−1.5±j0.866s^3+4s^2+6s+3=(s+1)(s^2+3s+3)\ \Longrightarrow\ s=-1,\ \ s=-1.5\pm j0.866
    5. The DC gain is b₀/a₀.

      G(0)=53≈1.667G(0)=\frac{5}{3}\approx 1.667
  4. This A has a repeated eigenvalue. Find e^{At} by inverse Laplace transform, then the zero-input response from x(0) = [1, 0]ᵀ, and evaluate it at t = 0.5 s.

    A=[01−4−4]A=\begin{bmatrix}0&1\\-4&-4\end{bmatrix}
    Show answer

    Answere^{At} = e^{−2t}[[1 + 2t, t], [−4t, 1 − 2t]], and x(0.5) = [0.736, −0.736]ᵀ.

    eAt=e−2t[1+2tt−4t1−2t]e^{At}=e^{-2t}\begin{bmatrix}1+2t&t\\-4t&1-2t\end{bmatrix}
    Show full solution

    Solution

    1. The characteristic polynomial is a perfect square, so λ = −2 twice.

      det⁡(sI−A)=s(s+4)+4=(s+2)2\det(sI-A)=s(s+4)+4=(s+2)^2
    2. Invert sI − A.

      (sI−A)−1=1(s+2)2[s+41−4s](sI-A)^{-1}=\frac{1}{(s+2)^2}\begin{bmatrix}s+4&1\\-4&s\end{bmatrix}
    3. Split the two entries that have s in the numerator.

      s+4(s+2)2=1s+2+2(s+2)2,s(s+2)2=1s+2−2(s+2)2\frac{s+4}{(s+2)^2}=\frac{1}{s+2}+\frac{2}{(s+2)^2},\qquad \frac{s}{(s+2)^2}=\frac{1}{s+2}-\frac{2}{(s+2)^2}
    4. Use the pair 1/(s + 2)² ↔ t·e^{−2t}. The factor t is the signature of a repeated eigenvalue. Check: at t = 0 the matrix is I.

      eAt=e−2t[1+2tt−4t1−2t]e^{At}=e^{-2t}\begin{bmatrix}1+2t&t\\-4t&1-2t\end{bmatrix}
    5. The response from x(0) = [1, 0]ᵀ is the first column. At t = 0.5 s the factor e^{−1} is 0.3679.

      x(t)=e−2t[1+2t−4t],x(0.5)=e−1[2−2]≈[0.736−0.736]x(t)=e^{-2t}\begin{bmatrix}1+2t\\-4t\end{bmatrix},\qquad x(0.5)=e^{-1}\begin{bmatrix}2\\-2\end{bmatrix}\approx\begin{bmatrix}0.736\\-0.736\end{bmatrix}
  5. Exam level. The model below starts at x(0) = [1, 0]ᵀ and receives a unit-step input. Find e^{At}, G(s) and y(t). Give y(0), y(1 s) and the final value.

    A=[01−5−2],B=[01],C=[10],D=0A=\begin{bmatrix}0&1\\-5&-2\end{bmatrix},\qquad B=\begin{bmatrix}0\\1\end{bmatrix},\qquad C=\begin{bmatrix}1&0\end{bmatrix},\qquad D=0
    Show answer

    Answery(t) = 1/5 + (4/5)e^{−t}(cos 2t + ½ sin 2t), with y(0) = 1, y(1) ≈ 0.211 and final value 0.2.

    y(t)=15+45e−t(cos⁡2t+12sin⁡2t)y(t)=\frac15+\frac45e^{-t}\left(\cos 2t+\tfrac12\sin 2t\right)
    Show full solution

    Solution

    1. Complete the square in the characteristic polynomial. The eigenvalues are complex: decay rate 1 s⁻¹ and oscillation at 2 rad/s.

      det⁡(sI−A)=s2+2s+5=(s+1)2+22 ⟹ λ=−1±j2\det(sI-A)=s^2+2s+5=(s+1)^2+2^2\ \Longrightarrow\ \lambda=-1\pm j2
    2. Invert sI − A and write each numerator in terms of (s + 1) so that the damped cosine and sine pairs apply.

      (sI−A)−1=1(s+1)2+4[(s+1)+11−5(s+1)−1](sI-A)^{-1}=\frac{1}{(s+1)^2+4}\begin{bmatrix}(s+1)+1&1\\-5&(s+1)-1\end{bmatrix}
    3. Transform with (s + 1)/((s + 1)² + 4) ↔ e^{−t}cos 2t and 2/((s + 1)² + 4) ↔ e^{−t}sin 2t.

      eAt=e−t[cos⁡2t+12sin⁡2t12sin⁡2t−52sin⁡2tcos⁡2t−12sin⁡2t]e^{At}=e^{-t}\begin{bmatrix}\cos 2t+\tfrac12\sin 2t&\tfrac12\sin 2t\\-\tfrac52\sin 2t&\cos 2t-\tfrac12\sin 2t\end{bmatrix}
    4. The transfer function is the top-right entry of the inverse. For the step, expand G(s)/s; the zero-state output settles at G(0) = 1/5.

      G(s)=1s2+2s+5,G(s)s=15[1s−s+2(s+1)2+4] ⟹ yzs(t)=15[1−e−t(cos⁡2t+12sin⁡2t)]G(s)=\frac{1}{s^2+2s+5},\qquad \frac{G(s)}{s}=\frac15\left[\frac1s-\frac{s+2}{(s+1)^2+4}\right]\ \Longrightarrow\ y_{zs}(t)=\frac15\left[1-e^{-t}\left(\cos 2t+\tfrac12\sin 2t\right)\right]
    5. Add the zero-input output, which is the top-left entry of e^{At}. Then y(0) = 1/5 + 4/5 = 1, y(1) = 0.2 + 0.8 × 0.3679 × 0.0385 ≈ 0.211, and y(∞) = 0.2.

      y(t)=e−t(cos⁡2t+12sin⁡2t)+yzs(t)=15+45e−t(cos⁡2t+12sin⁡2t)y(t)=e^{-t}\left(\cos 2t+\tfrac12\sin 2t\right)+y_{zs}(t)=\frac15+\frac45e^{-t}\left(\cos 2t+\tfrac12\sin 2t\right)

Mark each exercise after you check your answer. Marks stay in this browser.

KEY TERMS

Terms this lecture teaches

All terms in the glossary →
QUESTIONS STUDENTS ASK

Frequently asked questions

Why are the poles of the transfer function the eigenvalues of A?

Every entry of (sI − A)⁻¹ is a polynomial divided by det(sI − A), so G(s) = C(sI − A)⁻¹B + D shares that denominator. Its roots, the solutions of det(sI − A) = 0, are the eigenvalues of A. An eigenvalue is missing from G(s) only when a numerator factor cancels it.

Can I compute e^{At} by taking the exponential of each entry of A?

No. Exponentiating entry by entry is wrong for any A that is not diagonal. For A = [[0, 1], [−2, −3]] at t = 1 s, the top-left entry of e^{At} is about 0.600, not e⁰ = 1. Use the inverse Laplace transform of (sI − A)⁻¹, diagonalisation or Cayley–Hamilton, and check that the result equals I at t = 0.

Which method should I use to find e^{At} by hand?

The inverse Laplace transform of (sI − A)⁻¹ always works. Diagonalisation is quickest when the eigenvalues are distinct, but fails when a repeated eigenvalue lacks enough eigenvectors. The Cayley–Hamilton method writes e^{At} as α₀(t)I + α₁(t)A for two states and handles a repeated eigenvalue by differentiating its scalar equation.

What is the difference between the zero-input and zero-state response?

The zero-input response e^{At}x(0) is the motion caused by the initial state alone; the zero-state response is the convolution term the input produces from rest. By linearity the complete response is their sum, and the transfer function describes only the zero-state part.

Is there only one correct state-space model for a transfer function?

No. Any invertible change of coordinates x = T·x̄ gives new matrices with the same eigenvalues and the same transfer function. Controllable canonical, observable canonical and modal forms all describe the same input–output behaviour, so an answer in a different form can be fully correct.

Why the next lesson follows

The solution formula shows that the input reaches the state only through e^{A(t−τ)}B, and the state reaches the output only through Ce^{At}. The next chapter asks whether those two paths cover every state. Those questions are controllability and observability.

YOUR PROGRESS

Finished the reading and the exercises? Mark the lecture as done to tick it on the syllabus. Progress stays in this browser.

EIGENROOM · https://eigenroom.com/theory/state-solution · Updated 2026-11-15