From powers to continuous time
Definition 6.1.1
(Matrix exponential)
For a real square matrix and a real number , define
For any submultiplicative matrix norm, for the norm of the th term is at most , so the series converges absolutely. On each bounded interval of , the same bound with a fixed maximum for gives uniform convergence; the differentiated series has a similar uniform bound. Termwise differentiation is therefore justified, giving
Thus solves , .
A rotation from alternating powers
Let , so .
Theorem 6.1.2
(The exponential of is a rotation)
For every real ,
the counterclockwise rotation of the plane through angle .
Proof
(Even and odd powers of )
Even powers are and odd powers are . Separating the absolutely convergent series into even and odd terms gives
The first column is the image of , and the second is the image of . Their lengths are one and their dot product is zero; the determinant is . Hence this map preserves lengths and orientation, and sends to , so it rotates the plane counterclockwise through angle .
Proof X-ray
What the rotation proof actually uses
Two facts carry the argument. First, absolute convergence: it lets us regroup the terms of the series into even and odd subsequences and sum them separately, with no change in the total. Second, the single algebraic identity : it makes the powers of cycle with period four, so the coefficients of and are exactly the Maclaurin series of cosine and sine. Nothing else about is used; the conclusion comes from the power series, not from treating as an ordinary real number.
The identification as a rotation is a separate step, read from the columns: the first column is the image of , and the second is perpendicular to it with the same length. If we replaced by a matrix with , the signs would not alternate and the same splitting would give hyperbolic functions instead.
Worked examples
Worked example 6.1.3
(An initial-value problem with )
Take and . Then
At , the angle is and the solution is . Direct differentiation gives , and at it gives . In fact this solution is unique: if is another solution, then has zero derivative because commutes with its own power series; its value is . Also , as follows by multiplying the absolutely convergent series and collecting terms, so .
Worked example 6.1.4
(A nilpotent matrix gives a shear)
Let , so and the series stops: . For the solution of is . Check: and . Unlike the rotation, this map does not preserve lengths: has length .
Worked example 6.1.5
(A commuting pair gives a spiral)
Let and ; they commute, so because . For the solution of is , a counterclockwise spiral whose distance from the origin is . Check: since .
Multiplication needs attention
For commuting square matrices , multiplication and regrouping of the absolutely convergent series gives . Without commutation, that equality is not guaranteed.
Counterexample mode
Two nilpotent matrices need not combine as scalars
Claim: for all square matrices . Let and . Both squares vanish, so and , but
Meanwhile , so its even powers are and its odd powers are . Summing gives
The diagonal entries are equal, while the displayed product has diagonal entries and . Hence . Here . Repair: assume . This example shows that the scalar product rule cannot be used without checking a suitable condition; it does not claim commutation is necessary in every special case.
Exercises and solutions
Quick check
(Compute .)
Solution · Answer
.
Quick check
(For the initial-value problem with and , find .)
Solution · Answer
The angle is , so .
Quick check
(Why does preserve lengths?)
Solution · Answer
Its columns are orthonormal, so its transpose times itself is ; hence .
Quick check
(If , what is ?)
Solution · Answer
; every positive power of vanishes, so only the term of the series remains.
Quick check
(Find for .)
Solution · Answer
, so .