A Technical Note on Diagonalization Algorithm

By CHEUNG Lap Tin, Cliff (張立天)

§1 Introduction
Diagonalization algorithm is a fundamental technique in linear algebra that facilitates the
simplification of matrix operation by transforming a matrix into a diagonal form. This
transformation is particular useful in analyzing dynamic system, solving difference and
differential equations, and conducting stability analysis. In economics, diagonalization
algorithm plays a crucial role in both applied and theoretical context including but not
limited to solving linearized dynamic models (e.g. the Ramsey-Cass-Koopmans model),
which will be demonstrated later.

§2 Eigenvalues and eigenvectors, and Diagonalization algorithm

§2.1 Eigenvalues and eigenvectors

Definition 2.1.1
Let A be an 𝑛 × 𝑛 (square) matrix. An eigenvector 𝜉⃑ ≠ 0 lies along the same line as 𝐴𝜉⃑:
𝐴𝜉⃑ = 𝜆𝜉⃑, where 𝜆 is the eigenvalue.

(The vector 𝜉⃑ points in some direction. Upon applying the linear transformation A, a new
vector 𝐴𝜉⃑ is yielded. Given that 𝜉⃑ is an (eigen)vector, 𝐴𝜉⃑ points in the same or exactly
opposite direction as 𝜉⃑ . The transformation stretches, compresses, or reverses the
vector—but it does not rotate it off its line.)

Note:
1. Any scalar multiple is of an eigenvector is also an eigenvector of the system,
corresponding to the same eigenvalue. Any eigenvalue can be chosen for a given
eigenvalue,

2. If 𝐴 = 𝑐𝐼 for some 𝑐 ∈ ℝ, then every nonzero eigenvector 𝜉⃑ ∈ ℝ𝑛 is an eigenvector

of A with eigenvalue c, because 𝐴𝜉⃑ = (𝑐𝐼)𝜉⃑ = 𝑐(𝐼𝜉⃑) = 𝑐𝜉⃑.
To find the eigenvalues and eigenvectors of A is to solve the following system for 𝜆 ∈ ℝ
and nonzero vectors 𝜉⃑ such that

𝐴𝜉⃑ = 𝜆𝜉⃑ ⟺ 𝐴𝜉⃑ − 𝜆𝜉⃑ = ⃑0⃑ ⟺ (𝐴 − 𝜆𝐼)𝜉⃑ = ⃑0⃑

Note that (𝐴 − 𝜆𝐼)𝜉⃑ = 0 forms a system of n homogeneous linear equations. For

nontrivial solution, and equivalently, invertibility, det(𝐴 − 𝜆𝐼) = 0.

Theorem 2.1.1
Let A be an 𝑛 × 𝑛 matrix.

(1) The eigenvalues of A are the roots of the characteristic polynomial of A.

(2) The eigenvectors of A with eigenvalue 𝜆 are the nontrivial solution of the
homogeneous system (𝐴 − 𝜆𝐼)𝜉⃑ = ⃑0⃑.

Example 2.1.1
4 2
Find the eigenvalues and eigenvectors of 𝐴 = ( ).
1 3

4−𝜆 2
det(𝐴 − 𝜆𝐼) = | | = (4 − 𝜆)(3 − 𝜆) − 2, 𝜆1 = 5, 𝜆2 = 2
1 3−𝜆

For 𝜆1 = 5, its corresponding eigenvector is

−1 2 𝑥 0 1 −2 𝑥 0 2
( ) (𝑦 ) = ( ) ⟺ ( ) (𝑦) = ( ) , 𝜉⃑2 = ( )
1 −2 0 0 0 0 1

For 𝜆2 = 2, its corresponding eigenvector is

2 2 𝑥 0 1 1 𝑥 0 −1
( )( ) = ( ) ⟺ ( ) ( ) = ( ) , 𝜉⃑2 = ( )
1 1 𝑦 0 0 0 𝑦 0 1
§2.2 Diagonalization algorithm

Theorem 2.2.1
Let 𝜉⃑1 , … , 𝜉⃑𝑛 be the eigenvectors of an 𝑛 × 𝑛 matrix A with eigenvalues 𝜆1 , … , 𝜆𝑛 . Define
Λ = 𝑑𝑖𝑎𝑔(𝜆1 , … , 𝜆𝑛 ) and 𝑋 = (𝜉⃑1 , … , 𝜉⃑𝑛 ).

(1) The columns of 𝐴𝑋 = 𝑋Λ. Because:

𝐴𝑋 = 𝐴(𝜉⃑1 , … , 𝜉⃑𝑛 ) = (𝐴𝜉⃑1 , … , 𝐴𝜉⃑𝑛 ) = (𝜆1 𝜉⃑1 , … , 𝜆𝑛 𝜉⃑𝑛 )

𝜆1 0 ⋯ 0
0 𝜆2 ⋯ 0
𝑋Λ = (𝜉⃑1 , 𝜉⃑2 , … , 𝜉⃑𝑛 ) ( ) = (𝜆1 𝜉⃑1 , … , 𝜆𝑛 𝜉⃑𝑛 )
⋮ ⋮ ⋱ ⋮
0 0 ⋯ 𝜆𝑛

(2) If X is invertible, A is diagonalizable with 𝐴 = 𝑋ΛX −1 and equivalently Λ = X −1 𝐴𝑋.

Because: Since P is invertible, multiply both sides of 𝐴𝑋 = 𝑋Λ by 𝑋 −1

𝐴𝑋𝑋 −1 = 𝑋𝛬𝑋 −1

𝐴 = 𝑋𝛬𝑋 −1

(3) The eigenvector matrix X also diagonalizes all powers 𝐴𝑘 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑘 ∈ ℕ. Because:

𝐴 = 𝑋𝛬𝑋 −1 (𝑘 = 1)

Assume

𝐴𝑘 = 𝑋𝛬𝑘 𝑋 −1

𝐴𝑘+1 = 𝐴𝐴𝑘 = (𝑋𝛬𝑋 −1 )(𝑋𝛬𝑘 𝑋 −1 ) = 𝑋𝛬(𝑋 −1 𝑋)𝛬𝑘 𝑋 −1

By mathematical induction,

𝐴𝑘 = 𝑋𝛬𝑘 𝑋 −1 , 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑘 ∈ ℕ

Steps of diagonalization algorithm (distinct roots)

(1) Find the eigenvalues of the 𝑛 × 𝑛 matrix A.

(2) Find the eigenvectors of the matrix A corresponding to their eigenvalues.

(3) If X is invertible, then A is diagonalizable with 𝐴 = 𝑋ΛX −1 .

Example 2.2.1
4 1
Consider 𝐴 = ( ). Find 𝐴𝑘 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑘 ∈ ℕ.
0 3

4−𝜆 1
det(𝐴 − 𝜆𝐼) = | | = (4 − 𝜆)(3 − 𝜆), 𝜆1 = 4, 𝜆2 = 3
0 3−𝜆

For 𝜆1 = 4,

0 1 𝑥
( ⃑⃑, 𝜉⃑1 = (1)
) (𝑦) = 0
0 −1 0

For 𝜆2 = 3,

1 1 𝑥 −1
( ) (𝑦) = ⃑0⃑, 𝜉⃑2 = ( )
0 0 1

Therefore,

4 0 1 −1
𝛬=( ),𝑋 = ( )
0 3 0 1

Find 𝑋 −1 then,

1 1 1 1 1
𝑋 −1 = ( )=( )
1 0 1 0 1

Since 𝐴𝑘 = 𝑋𝛬𝑘 𝑋 −1,

1 −1 4 0 𝑘 1 1 1 −1 4𝑘 0 ) (1 1
𝐴𝑘 = ( )( ) ( )=( )( )
0 1 0 3 0 1 0 1 0 3𝑘 0 1

𝑘
𝐴𝑘 = (4 −3𝑘 ) (1 1
) = (4
𝑘
4𝑘 − 3𝑘 )
0 3𝑘 0 1 0 3𝑘
Example 2.2.2 (Stability in the Ramsey-Cass-Koopmans model)
Consider the following canonical Ramsey-Cass-Koopmans problem:

∞
max ∫ 𝑒 −(𝜌−𝑛)𝑡 𝑢(𝑐(𝑡))𝑑𝑡
[𝑐(𝑡),𝑘(𝑡)]∞
𝑡=0 0

𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑘̇ (𝑡) = 𝑓(𝑘(𝑡)) − 𝑐(𝑡) − (𝑛 + 𝛿)𝑘(𝑡),
lim 𝜇(𝑡)𝑘(𝑡),
𝑡→∞
𝑤𝑖𝑡ℎ 𝑘(0) 𝑔𝑖𝑣𝑒𝑛

(Steps involved in solving the problem are omitted. Please refer to another note of mine
demonstrating the use of the Hamiltonian function.)

Assume the production function is Harrod-neutral and Cobb-Douglas and the utility
function is isoelastic with 𝜃 ≠ 1. The solution to the RCK problem is:

𝑐(𝑡)
𝑐̇ (𝑡) = (𝛼𝑘(𝑡)𝛼−1 − 𝑛 − 𝛿 − 𝜌)
𝜃

The other dynamics of the model is represented by the equation of motion for capital

𝑘̇ (𝑡) = 𝑘(𝑡)𝛼 − 𝑐(𝑡) − (𝑛 + 𝛿)𝑘(𝑡)

Then, find the steady state of k and c

1
∗
𝜌 + 𝑛 + 𝛿 𝛼−1
𝑘 =( )
𝛼

𝑐 ∗ = (𝑘 ∗ )𝛼 − (𝑛 + 𝛿)𝑘 ∗

Define the deviation of k and c from the steady state as 𝑘̃ = 𝑘 − 𝑘 ∗ and 𝑐̃ = 𝑐 − 𝑐 ∗ . Apply
the first-order Taylor linear approximation such that the pair of nonlinear differential
equation can be written as

𝑘̇ (𝑡) 𝑘(𝑡) − 𝑘 ∗
( ) ≈ 𝐽( )
𝑐̇ (𝑡) 𝑐(𝑡) − 𝑐 ∗

𝜕𝑘̇ 𝜕𝑘̇

where 𝐽 = (𝜕𝑘
𝜕𝑐̇
𝜕𝑐
𝜕𝑐̇
)| is the Jacobian matrix
𝜕𝑘 𝜕𝑐 (𝑘=𝑘 ∗,𝑐=𝑐 ∗ )
Substitute the corresponding values to get

𝛼(𝑘 ∗ )𝛼−1 − (𝑛 + 𝛿) −1
𝑘̇ (𝑡) 𝑘(𝑡) − 𝑘 ∗
( ) ≈ ( 𝑐 ∗ 𝛼(𝛼 − 1) ∗ 𝛼−2 1 )( ∗)
𝑐̇ (𝑡) (𝑘 ) (𝛼(𝑘 ∗ )𝛼−1 − 𝑛 − 𝛿 − 𝜌) 𝑐(𝑡) − 𝑐
𝜃 𝜃

1
Since 𝛼𝑘(𝑡)𝛼−1 − 𝑛 − 𝛿 − 𝜌 = 0 and 𝜃 (𝛼𝑘(𝑡)𝛼−1 − 𝑛 − 𝛿 − 𝜌) = 0 at the steady state
𝜌 −1
𝑘̇ (𝑡) ∗ 𝑘(𝑡) − 𝑘 ∗
( ) ≈ (𝑐 𝛼(𝛼 − 1) ∗ 𝛼−2 )( )
𝑐̇ (𝑡) (𝑘 ) 0 𝑐(𝑡) − 𝑐 ∗
𝜃

𝑐 ∗ 𝛼(𝛼−1)
Simplify the system by letting 𝛾 = (𝑘 ∗ )𝛼−2
𝜃

𝑘̇ (𝑡) 𝜌 −1 𝑘(𝑡) − 𝑘 ∗
( )≈( )( )
𝑐̇ (𝑡) 𝛾 0 𝑐(𝑡) − 𝑐 ∗

Find the eigenvalues of the Jacobian

𝜌−𝜆 −1 𝜌 + √𝜌2 − 4𝛾 𝜌 − √𝜌2 − 4𝛾

det(𝐴 − 𝜆𝐼) = | | = 𝜆2 − 𝜌𝜆 + 𝛾, 𝜆1 = , 𝜆2 =
𝛾 −𝜆 2 2

𝑐 ∗ 𝛼(𝛼−1)
Given that 𝛼 ∈ (0, 1), 𝛾 = (𝑘 ∗ )𝛼−2 < 0, 𝜌2 − 4𝛾 > 0. 𝜆1 > 0.
𝜃

For 𝜆2 , −4𝛾 > 0, 𝜌2 − 4𝛾 > 𝜌2 , √𝜌2 − 4𝛾 > 𝜌, 𝜌 − √𝜌2 − 4𝛾 < 0, so 𝜆2 < 0.

(The eigenvectors are not solved for their negligibility in the problem.)

Solving the system of linear ODEs, we have

𝑘(𝑡) − 𝑘 ∗ 𝑡𝐽 𝑘(0) − 𝑘
∗
𝑡𝐷 −1 𝑘(0) − 𝑘
∗
( ) ≈ 𝑒 ( ) = 𝑋𝑒 𝑋 ( ) = 𝐶1 𝑒 𝜆1𝑡 𝜉⃑1 + 𝐶2 𝑒 𝜆2𝑡 𝜉⃑2
𝑐(𝑡) − 𝑐 ∗ 𝑐(0) − 𝑐 ∗ 𝑐(0) − 𝑐 ∗

For 𝜆1 > 0, lim 𝑒 𝜆1𝑡 = ∞ such that the trajectory is unstable.

𝑡→∞
For 𝜆2 < 0, lim 𝑒 𝜆2𝑡 = 0 such that the trajectory is asymptotically stable.
𝑡→∞
The solution diverges from the steady state unless C1 = 0, because the component along
𝜉⃑1 grows without bound due to 𝜆1 > 0.

To make the system converge to the steady state, we must set C1 = 0 so the unstable part
disappears. The trajectory is known as the saddle path.
Example 2.2.3 (Numerical solution of the Real Business Cycle (RBC) model)
Consider the following social planner’s problem:

max 𝔼1 [∑ 𝛽𝑡−1 𝑢(𝑐𝑡 , 1 − ℎ𝑡 )]

{𝑐𝑡 ,ℎ𝑡 }∞
𝑡=1
𝑡=1

𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑐𝑡 + 𝑖𝑡 = 𝑧𝑡 𝑓(𝑘𝑡 , ℎ𝑡 ) ≡ 𝑦𝑡 ,
𝑘𝑡+1 = 𝑘𝑡 (1 − 𝛿) + 𝑖𝑡 ,
𝜌
𝑧𝑡+1 = 𝑧𝑡 𝜀𝑡+1 ,
𝑤𝑖𝑡ℎ 𝑘1 𝑔𝑖𝑣𝑒𝑛,

𝑤ℎ𝑒𝑟𝑒
𝛽 ∈ (0, 1) 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑎𝑔𝑒𝑛𝑡 ′ 𝑠 𝑠𝑢𝑏𝑗𝑒𝑐𝑡𝑖𝑣𝑒 𝑑𝑖𝑠𝑐𝑜𝑢𝑛𝑡 𝑟𝑎𝑡𝑒,
(1 − ℎ𝑡 ) 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑙𝑒𝑖𝑠𝑢𝑟𝑒 𝑤𝑖𝑡ℎ 𝑡ℎ𝑒 𝑒𝑛𝑑𝑜𝑤𝑚𝑒𝑛𝑡 𝑜𝑓 𝑡𝑖𝑚𝑒 𝑛𝑜𝑟𝑚𝑎𝑙𝑖𝑧𝑒𝑑 𝑡𝑜 1,
𝑖𝑡 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑖𝑛𝑣𝑒𝑠𝑡𝑚𝑒𝑛𝑡,
𝛿 ∈ (0, 1) 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑟𝑎𝑡𝑒 𝑜𝑓 𝑐𝑎𝑝𝑖𝑡𝑎𝑙 𝑑𝑒𝑝𝑟𝑒𝑐𝑖𝑎𝑡𝑖𝑜𝑛,
𝑧 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑒𝑥𝑜𝑔𝑒𝑛𝑜𝑢𝑠 𝑡𝑒𝑐ℎ𝑛𝑜𝑙𝑜𝑔𝑦 𝑓𝑜𝑙𝑙𝑜𝑤𝑖𝑛𝑔 𝑎𝑛 𝑎𝑢𝑡𝑜𝑟𝑒𝑔𝑟𝑒𝑠𝑠𝑖𝑣𝑒 𝑝𝑟𝑜𝑐𝑒𝑠𝑠,
𝜌 ∈ (0, 1) 𝑑𝑒𝑛𝑜𝑡𝑒𝑠 𝑡ℎ𝑒 𝑎𝑢𝑡𝑜𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛 𝑝𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟.

Set up the Bellman equation

𝑉(𝑘𝑡 , 𝑧𝑡 ) = max {𝑢(𝑐𝑡 , 1 − ℎ𝑡 ) + 𝛽𝔼𝑡 [𝑉(𝑘𝑡+1 , 𝑧𝑡+1 )]}

𝑐𝑡 ,ℎ𝑡 ,𝑘𝑡+1

𝑠𝑢𝑏𝑗𝑒𝑐𝑡 𝑡𝑜
𝑐𝑡 + 𝑘𝑡+1 = 𝑧𝑡 𝑓(𝑘𝑡 , ℎ𝑡 ) + 𝑘𝑡 (1 − 𝛿)
𝜌
𝑧𝑡+1 = 𝑧𝑡 𝜀𝑡+1

(Steps involved in solving the problem are omitted. Please refer to another note of mine
demonstrating the use of the Bellman equation.)

The solution must follow the following necessary conditions and resource constraint.

𝜕𝑉𝑡
From =0
𝜕ℎ𝑡

𝑢1−ℎ𝑡 = 𝑢𝑐𝑡 𝑧𝑡 𝑓ℎ𝑡 (1)

𝜕𝑉𝑡
From 𝜕𝑘 =0
𝑡+1

𝑢𝑐𝑡 = 𝛽𝔼𝑡 [𝑉𝑘𝑡+1 ]

Also

𝑉𝑘𝑡+1 = 𝑢𝑐𝑡+1 (𝑧𝑡+1 𝑓𝑘𝑡+1 + (1 − 𝛿))

Therefore

𝑢𝑐𝑡 = 𝛽𝔼𝑡 [𝑢𝑐𝑡+1 (𝑧𝑡+1 𝑓𝑘𝑡+1 + (1 − 𝛿))] (2)

The resource constraint is represented by

𝑘𝑡+1 = 𝑧𝑡 𝑓(𝑘𝑡 , ℎ𝑡 ) + 𝑘𝑡 (1 − 𝛿) − 𝑐𝑡 (3)

Subscripts on a function denote partial derivatives with respect to the corresponding

𝜕𝑢 𝜕𝑓
argument, e.g., 𝑢𝑥 = 𝜕𝑥 𝑎𝑛𝑑 𝑓𝑦 = 𝜕𝑦 .

Assume the following functional forms for preference and production

𝑢(𝑐𝑡 , 1 − ℎ𝑡 ) = ln 𝑐𝑡 + 𝐴(1 − ℎ𝑡 ) (𝑝𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒)

𝑓(𝑘𝑡 , ℎ𝑡 ) = 𝑘𝑡𝛼 ℎ1−𝛼

𝑡 (𝑝𝑟𝑜𝑑𝑢𝑐𝑡𝑖𝑜𝑛)

The steady-state equilibrium for the economy is one in which the technology shock is
assumed to be constant, so 𝑧𝑡 = 1 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑡.

With the specified functional forms for preference and production, by (1)

1−𝛼
𝑐𝑡 = ( ) 𝑧𝑡 𝑘𝑡𝛼 ℎ𝑡−𝛼 (4)
𝐴

Also, by (2)

𝑐𝑡−1 = 𝛽𝔼𝑡 [𝑐𝑡+1

−1 𝛼−1 1−𝛼
(𝛼𝑧𝑡+1 𝑘𝑡+1 ℎ𝑡+1 + (1 − 𝛿))] (5)

Linearize (4) around the steady state to get

1 − 𝛼 ∗ 𝛼−1 ∗ −𝛼 1 − 𝛼 ∗ 𝛼 ∗ −𝑎−1
(𝑐𝑡 − 𝑐 ∗ ) = 𝛼 ( )𝑘 ℎ (𝑘𝑡 − 𝑘 ∗ ) − 𝛼 ( )𝑘 ℎ (ℎ𝑡 − ℎ∗ )
𝐴 𝐴
1 − 𝛼 ∗ 𝛼 ∗ −𝛼
+( ) 𝑘 ℎ (𝑧𝑡 − 𝑧 ∗ )
𝐴
 1 − 𝛼 ∗ 𝛼 ∗ −𝛼 𝑘𝑡 − 𝑘 ∗ 1 − 𝛼 ∗ 𝛼 ∗ −𝑎 ℎ𝑡 − ℎ ∗
= 𝛼( )𝑘 ℎ ( ) − 𝛼 ( )𝑘 ℎ ( )
𝐴 𝑘∗ 𝐴 ℎ∗
1 − 𝛼 ∗ 𝛼 ∗ −𝛼 𝑧𝑡 − 𝑧 ∗
+( )𝑘 ℎ ( )
𝐴 𝑧∗

Defining variables with tilde on top as the percentage deviation from the steady state,
we have

𝑐̃𝑡 = 𝛼𝑘̃𝑡 − 𝛼ℎ̃𝑡 + 𝑧̃𝑡 (6)

Linearize (2) to get

−𝑐 ∗ −1 𝑐̃𝑡 = −𝛽𝑐 ∗ −1 [𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿)]𝔼𝑡 (𝑐̃𝑡+1 )

+ 𝛽𝑐 ∗ −1 𝛼(𝛼 − 1)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (𝑘̃𝑡+1 )
+ 𝛽𝑐 ∗ −1 𝛼(𝛼 − 1)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (ℎ̃𝑡+1 ) + 𝛽𝑐 ∗ −1 𝛼𝑘 ∗ 𝛼−1 ℎ ∗1−𝛼 𝔼𝑡 (𝑧̃𝑡+1 )

−𝑐̃𝑡 = −𝛽[𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿)]𝔼𝑡 (𝑐̃𝑡+1 ) + 𝛽𝛼(𝛼 − 1)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (𝑘̃𝑡+1 )
+ 𝛽𝛼(1 − 𝛼)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (ℎ̃𝑡+1 ) + 𝛽𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (𝑧̃𝑡+1 )

Under the steady state, (5) is transformed into

𝑐 ∗ −1 = 𝛽𝑐 ∗ −1 (𝛼𝑧𝑘 𝛼−1 ℎ1−𝛼 + (1 − 𝛿))

1 = 𝛽 (𝛼𝑧 ∗ 𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿))

1 = 𝛽 (𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿))

Substitute it into the earlier linearized equation to get

−𝑐̃𝑡 = −𝔼𝑡 (𝑐̃𝑡+1 ) + 𝛽𝛼(𝛼 − 1)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (𝑘̃𝑡+1 ) + 𝛽𝛼(1 − 𝛼)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (ℎ̃𝑡+1 ) +
𝛽𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝔼𝑡 (𝑧̃𝑡+1 ) (7)

Linearize the resource constraint in similar manner to get

𝑘̃𝑡+1 = [𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿)]𝑘̃𝑡 + [(1 − 𝛼)𝑘 ∗ 𝛼−1 ℎ∗ (1−𝛼) ]ℎ̃𝑡 + (𝑘 ∗ 𝛼−1 ℎ∗ (1−𝛼) )𝑧̃𝑡 −
𝑐∗
(𝑘 ∗) 𝑐̃𝑡 (8)
Linearize the autoregressive process for technology shock to get

𝑧̃𝑡+1 = 𝜌𝑧̃𝑡 + 𝜀̃𝑡+1 (9)

Take the expectation at time t to get

𝔼𝑡 (𝑧̃𝑡+1 ) = 𝔼𝑡 (𝜌𝑧̃𝑡 + 𝜀̃𝑡+1 ) = 𝔼𝑡 (𝜌𝑧̃𝑡 ) + 𝔼𝑡 (𝜀̃𝑡+1 )

Since the shock is a mean-zero innovation uncorrelated to time t

𝔼𝑡 (𝑧̃𝑡+1 ) = 𝜌𝑧̃𝑡 (10)

Equations (6), (7), (8) and (10) can be written as a vector expectational difference
equation. Let

𝑐̃𝑡
𝑘̃𝑡
𝑢
⃑⃑𝑡 =
ℎ̃𝑡
( 𝑧̃𝑡 )

The linear system of difference equations can be expressed by

⃑⃑𝑡 = 𝐵𝔼𝑡 (𝑢
𝐴𝑢 ⃑⃑𝑡+1 ) (11)

where

1 −𝛼 𝛼 −1
−1 0 0 0
𝐴= 𝑐∗
− ∗ 𝛼𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 + (1 − 𝛿) (1 − 𝛼)𝑘 ∗ 𝛼−1 ℎ∗ (1−𝛼) 𝑘 ∗ 𝛼−1 ℎ∗
(1−𝛼)
𝑘
( 0 0 0 𝜌 )

0 0 0 0
−1 𝛽𝛼(𝛼 − 1)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝛽𝛼(1 − 𝛼)𝑘 ∗ 𝛼−1 ℎ∗1−𝛼 𝛽𝛼𝑘 ∗ 𝛼−1 ∗1−𝛼
ℎ
𝐵=( )
0 1 0 0
0 0 0 1

Multiply both sides of (11) by 𝐴−1 to get

⃑⃑𝑡 = 𝐴−1 𝐵𝔼𝑡 (𝑢

𝑢 ⃑⃑𝑡+1 ) (12)
Using diagonalization algorithm, 𝐴−1 𝐵 can be decomposed into

𝐴−1 𝐵 = 𝑋Λ𝑋 −1

where 𝑋 = (𝜉⃑1 , … , 𝜉⃑4 ) and Λ = 𝑑𝑖𝑎𝑔(𝜆1 , … , 𝜆4 ). Such that (12) becomes

⃑⃑𝑡 = 𝑋Λ𝑋 −1 𝔼𝑡 (𝑢
𝑢 ⃑⃑𝑡+1 )

Multiply both side of the expression by 𝑋 −1 to get

𝑋 −1 𝑢
⃑⃑𝑡 = Λ𝑋 −1 𝔼𝑡 (𝑢
⃑⃑𝑡+1 ) = Λ𝔼𝑡 (𝑋 −1 𝑢
⃑⃑𝑡+1 )

Define 𝑑𝑡 ≡ 𝑋 −1 𝑢
⃑⃑𝑡 and 𝑑𝑡+1 = 𝑋 −1 𝑢
⃑⃑𝑡+1

𝑑𝑖,𝑡 = 𝜆𝑖 𝔼𝑡 (𝑑𝑖,𝑡+1 ); 𝑖 = 1, 2, 3, 4 (13)

Since the equation must hold every period, upon recursively substituting the expressions
forward for T period

𝑑𝑖,𝑡 = 𝜆𝑇𝑖 𝔼𝑡 (𝑑𝑖,𝑡+1 ); 𝑖 = 1, 2, 3, 4

Now, we analyze the stability of the system by referring to the value of the eigenvalues

The first rows of the matrices A and B are determined by the intratemporal efficiency
condition ( 𝑢1−ℎ𝑡 = 𝑢𝑐𝑡 𝑧𝑡 𝑓ℎ𝑡 ), there is no dynamics (it is not a difference equation
technically). Since one row is zero, the determinant of the matrix 𝐴−1 𝐵 is zero, which
means it is singular. Therefore, at least one eigenvalue is zero. It is static and does not give
the solution of the dynamic problem.

⃑⃑𝑡 = 𝐵𝔼𝑡 (𝑢
In the system 𝐴𝑢 ⃑⃑𝑡+1 ), the last row gives

𝜌𝑧̃𝑡 = 𝔼𝑡 (𝑧̃𝑡+1 )

𝑧̃𝑡 = 𝜌−1 𝔼𝑡 (𝑧̃𝑡+1 )

in which 𝜌 −1 corresponds to 𝜆𝑖 in (13). Therefore, one of the eigenvalues is 𝜌−1 , which is

greater than 1 and (13) is unstable.

Therefore, we seek for the eigenvalue of which its (absolute) value is in between 0 and 1
in order to solve for the rational expectation solution in calibration.
Reference
Barro, R. J., & Sala-i-Martin, X. (2004). Economic growth (2nd ed.). MIT Press.

Hartley, J. E., Hoover, K. D., & Salyer, K. D. (1998). Real business cycles : a reader (1st
edition). Routledge.

Strang, G. (2016). Introduction to linear algebra (Fifth edition.). Cambridge Press.