Why two numbers help, and where they do not
For the autonomous system in the real plane, the solution is . Eigenvalues suggest whether trajectories approach or leave the origin, but an eigenvalue list alone does not show how to construct every solution when eigenvectors are missing. The trace and determinant of a real matrix organize the eigenvalues, while a simple matrix identity produces a formula that also works at repeated roots.
This section concerns the linear system itself. A nonlinear system can behave differently when its linearization has eigenvalues on the imaginary axis; no nonlinear conclusion follows from the borderline linear classification alone.
One exponential formula for all three discriminants
Definition 4.3.1
(Trace–determinant data)
For a real matrix let , , and put
The number is a quarter of the discriminant of the characteristic polynomial .
The three cases , , will not need three separate derivations.
Theorem 4.3.2
(Exponential of a real matrix)
With this notation , and for every real
where the real functions are
Proof
(Even and odd powers of )
Cayley–Hamilton gives and hence . Since commutes with , . Because and , the series for splits into the even series and the odd series , which converge for every .
For instance, when the series terminates after the linear term in . Formula (1) can also be verified by differentiation: , , and its value at zero is . Uniqueness for the linear initial value problem then identifies it with the fundamental matrix.
Proof X-ray
Why the repeated-root formula has no division by zero
The key move is the shift by the mean eigenvalue . Because commutes with , the exponential factors as , and Cayley–Hamilton, rewritten around , says that has a scalar square. Every power of is then a scalar multiple of or of , so two scalar series do all the work.
Nothing in the argument divides by the eigenvalue gap . At the series stops at , whether is zero or not, and the table is continuous in : and both tend to . A formula with must instead be replaced by this limit. The differentiation check closes the argument independently of the series, through uniqueness.
The eigenvalues over are ; for negative interpret the square root as .
Real roots, complex roots, and the boundary
When , there are two distinct real eigenvalues , with independent real eigenvectors. In the eigenvector coordinates the components are .
- If , the eigenvalues have opposite signs: the origin is a saddle. The stable and unstable eigendirections give trajectories converging to zero in forward and backward time respectively; generic nonzero trajectories escape in forward time.
- If , both eigenvalues have the same sign, determined by : negative trace gives a stable node, positive trace an unstable node. In the stable case every solution tends to zero exponentially.
- If , one eigenvalue is zero and the other is . There is a line of equilibria. If , trajectories converge to points on that line; if , the origin is unstable. A zero eigenvalue must not be classified as a decaying direction.
When , write . Then . For any nonzero real vector , the vectors and are independent: otherwise would have a real eigenvalue, contradicting its square being . In this basis is the matrix . Thus (1) is a rotation in these coordinates multiplied by . Negative gives a stable spiral, positive an unstable spiral, and a center: every nonzero orbit is periodic. The coordinate change need not preserve Euclidean circles, so center trajectories in the original coordinates may be ellipses.
When , the repeated eigenvalue is and (1) becomes . If , every vector is an eigenvector. If , its nonzero nilpotent part introduces a linear factor and there is only one independent eigenvector. For , all trajectories nevertheless decay exponentially: for example, for . For , a nonzero eigenvector grows exponentially. At , makes every point an equilibrium, while gives unbounded trajectories for initial states outside ; the origin is unstable. These borderline cases cannot be settled by looking only at the repeated eigenvalue's real part.
The case analysis above already contains the classical stability test.
Theorem 4.3.3
(Asymptotic stability from trace and determinant)
For a real matrix , every solution of tends to as if and only if and .
Proof
(Reading the decay rate from formula (1))
Suppose and , so . If , then , so ; since and , both coefficients in (1) are bounded by a constant times . If , . If , are bounded and . Hence .
Conversely, if , then , so is a real eigenvalue with a real eigenvector , and does not tend to . If and , then : for the same eigenvector argument applies, and for the solution is times a rotation in the basis , so its coordinates in that basis have length times a positive constant and do not tend to .
The boundary cases of the theorem are exactly the ones where the classification needs care, as the next two examples show.
Worked example 4.3.4
(A center whose orbits are ellipses)
Let . Then , , , , and . Formula (1) gives
From the solution is , with period ; indeed and . It satisfies , so the orbit is an ellipse with semi-axes and , although the origin is a center.
Common mistake
(Expecting circles at a center)
When and , formula (1) is a rotation only in the basis , . In standard coordinates the orbits are ellipses in general. They are circles exactly when preserves length, that is, when is skew-symmetric.
Worked example 4.3.5
(A zero determinant and a line of equilibria)
Let . Then , , , and . Since and , formula (1) gives
Directly, is constant and gives . The equilibria form the line , which is , and the solution from tends to , not to the origin. Here but : the origin is stable, since , but not asymptotically stable.
Common mistake
(Treating a zero eigenvalue as a decaying direction)
When , one eigenvalue is and its eigenvectors are equilibria. A negative trace makes the other direction decay, but solutions then approach a point of the equilibrium line, as in the last example, not the origin.
The explorer below moves a real matrix through the trace–determinant plane and evaluates formula (1) for , with the exact check , the class and stability of the origin, and the orbits of , including the borderline cases where only settles the answer.
Worked example: a stable system with transient growth
Worked example 4.3.6
(Transient growth at a repeated root)
Take
Here , and , so
The solution tends to zero, but at its Euclidean norm is . Asymptotic stability is a statement about sufficiently small initial conditions remaining controlled and eventually tending to zero; it does not require every individual norm to decrease at every time.
Counterexample mode
The eigenvalue list does not determine the full flow
Both and have the repeated eigenvalue . Yet , whereas . They have the same asymptotic stability but different trajectories and transient behavior. The distinction is the nilpotent part, not another eigenvalue.
Practice
Quick check
(For , calculate and classify the origin.)
Solution · Answer
Here , , and . Hence . Eigenvalues give a saddle.
Quick check
(For , solve the system from and determine the orbit period.)
Solution · Answer
Here , , and (1) gives . The orbit has period ; the origin is a center.
Quick check
(For , find all equilibria and determine stability of the origin.)
Solution · Answer
and . Equilibria form the line . Any nonzero initial , however small, makes the first coordinate unbounded, so the origin is unstable.
Quick check
(Does alone imply asymptotic stability? Give a concrete counterexample.)
Solution · Answer
No. For , the determinant is but grows for every nonzero initial state. The trace and the type of the roots also matter.