In 7.2 Linear Systems & the Eigenvalue Method we solved when the eigenvalues were real and distinct. This lesson asks a simpler question: what does the picture look like, and can we know it before solving anything? For a matrix the answer fits on one page of axes, because two numbers, the trace and the determinant , decide the whole picture. The payoff for simulation is concrete: the particle paths of a steady 2D flow near a critical point are exactly such a system, and whether a vortex stays a vortex is a statement about the line . We also meet the case this course cares about most: explicit Euler on a pure rotation does not rotate, it spirals out.
Transcript
- For x prime equals A x, the picture of all trajectories, with arrows for the direction of time, is the phase portrait. Two numbers decide it: the trace T and the determinant D. We move one point around the T D plane and watch the portrait change.
- Start with the matrix with rows 1, 2 and 3, 2. Its eigenvalues are 4 and minus 1, real with opposite signs, which a negative determinant guarantees: a saddle. Only the stable line approaches the origin; every other curve is bent away along the unstable one.
- Same-sign real eigenvalues give a node. Rows minus 4, 2 and 1, minus 3 have T minus 7, D 10, T squared minus 4 D 9, so lambda is minus 2 and minus 5: stable. Curves enter tangent to the slow line. Rows 4, minus 2 and 1, 1 give 3 and 2: unstable, time reversed.
- A complex pair with nonzero real part spirals. Rows minus 0.2, minus 1 and 1, minus 0.2 give lambda equal to minus 0.2 plus or minus i. Each lap multiplies the radius by 0.28461: 28.5 percent is left. A 21 is positive, so it turns counterclockwise.
- With the real part zero the spiral closes into ellipses: a center, lambda plus or minus 2 i, period pi, factor exactly 1.00000 per turn. Nearby starts stay nearby: stable, but not asymptotically. Here a 21 is minus 4, so it turns clockwise.
- Repeated eigenvalues sit on the parabola. Minus I is a star node: every vector is an eigenvector, so every trajectory is a straight ray. Rows minus 2, 1 and 0, minus 2 have one eigenline only: a degenerate node, every curve tangent to it. Why is the next lesson.
- Expanding the determinant gives lambda squared minus T lambda plus D equals 0, so lambda is T plus or minus the square root of T squared minus 4 D, over 2. Negative D is a saddle; positive D gives nodes below the parabola, spirals above, a center on T equals 0. Stable exactly when T is negative and D positive.
- The oscillator y double prime plus c y prime plus k y equals 0, with k equal to 4, has T equal to minus c and D equal to 4. Increase c. At 0, a center. At 2, a spiral, minus 1 plus or minus 1.7321 i. At 4, on the parabola, minus 2 twice. At 5, a node.
- A center needs T exactly 0. Add 0.01 to a 11: T is 0.01, the real part 0.005, and each turn multiplies the radius by 1.01583, a slow outward spiral. Minus 4.01 in a 21 keeps T at 0, lambda plus or minus 2.0025 i: still a center, but a thin slice of perturbations.
- The parabola is a hinge. Perturb rows minus 2, 1 and 0, minus 2 by epsilon in a 21: T squared minus 4 D is 4 epsilon. Epsilon plus 0.01 gives minus 2.1 and minus 1.9, a node; minus 0.01 gives minus 2 plus or minus 0.1 i, a spiral.
- Explicit Euler multiplies a center's amplitude by the square root of 1 plus h squared omega squared each step. At h 0.1, 31 steps, the gain is 1.8366; at 0.05, 63 steps, 1.3681; ten laps, 22.7463. RK4 stays near 0.99998629. Euler expands area: det of I plus h A is 1.04.
- Near a critical point of a steady 2D flow, A is the Jacobian, and its trace is the divergence. Incompressible means T equals 0: only saddles, like u equals x, minus y, and centers, like minus 2 y, 2 x. From memory, sources and sinks need divergence.
- To sum up. D and T squared minus 4 D give the picture; stable means T negative, D positive. The center and the parabola are fragile. Euler turns a center into an outward spiral. Incompressible flows have only saddles and centers. Next: complex and repeated eigenvalues.
Five pictures from the eigenvalues
A phase portrait is the plane of states drawn with many trajectories, each with an arrow for the direction of time. The origin is the only equilibrium when is invertible. The eigenvalues (from 7.2 Linear Systems & the Eigenvalue Method) decide everything. Concrete matrices, with eigenvalues computed in block A of the lab cell below and shown by the sim:
| Type | Matrix | |
|---|---|---|
| saddle | ||
| stable node | ||
| unstable node | ||
| stable spiral | ||
| center |
Saddle (real, opposite signs). The two eigenlines carry the two straight-line solutions and . One is pulled in, the other pushed out, so every trajectory off the stable line is bent away along the unstable one. Only the stable line approaches the origin; a point on it that is nudged off drifts away, which is the knife-edge of the last lesson.
Node (real, same sign). Both modes grow, or both decay. In the stable node the mode dies first, so every curve bends onto the slow eigenline (, direction ) and enters the origin tangent to it. In the unstable node, running time backwards does the same, so curves leave tangent to the slow () line, and later turn to follow the fast one.
Spiral or focus (complex pair, ). A solution is times something that rotates at angular rate . We take that on trust here, since turning the complex solution into a real one is the work of 7.4 Complex & Repeated Eigenvalues, and we check it by measurement. If the radius shrinks while it turns, so the curve spirals in; if , out.
Center (pure imaginary). : the amplitude never changes, so trajectories are closed loops around the origin, ellipses for our matrix. It is stable (nearby starts stay nearby) but not asymptotically stable.
Which way does it turn? At the point the velocity is . A complex pair needs , which forces , so the point at must be moving sideways around the origin, and the sign of says whether it moves up or down. So means counterclockwise. The stable spiral above has , counterclockwise; the center has , clockwise. The sim’s readout prints the sign of and, as a check, the sign of .
How fast does a spiral close? After one full turn, of duration , the rotating part is back where it started and only the factor is left:
For this is : each lap around the origin leaves 28.5 percent of the radius. The measured value is (block B, RK4 over one period from , and the sim’s readout). Because depends only on the ratio , a small ratio means close to 1 and many laps before the radius shrinks much.
Degenerate and star nodes
When the eigenvalue is repeated, , there are two possibilities. (the sim’s ) is a star node: every vector is an eigenvector, so every trajectory is a straight ray into the origin. has the same repeated eigenvalue but only one eigenline, and is a degenerate (improper) node: every curve is tangent to that single line. These sit on the parabola in the next section. Why there is only one eigenvector, and the extra solution that replaces the missing one, is the generalised eigenvector story of 7.4 Complex & Repeated Eigenvalues; here we only classify and look.
Trace and determinant: the whole classification in one plane
Expand for :
So the eigenvalues are
and, comparing with , and . Now read the pictures off the sign of and of .
- : the product is negative, so the roots are real with opposite signs: saddle (and is automatic).
- and : real roots of the same sign, the sign of : node, stable if .
- and : a complex pair with real part : spiral, stable if .
- and : : center.
- : repeated root : the parabola, with star or degenerate nodes on it.
- : a root , a whole line of equilibria rather than an isolated one.
Stability. The origin is asymptotically stable when both . For real roots that means the same sign, , and a negative sum, . For a complex pair and holds automatically. So in both cases: stable if and only if and . The test needs no eigenvalues at all, which is why it appears in every control textbook.
Worked examples
Example 1: a full sketch from and . Take . Then , and . Since and it is a node, and makes it stable. The eigenvalues are and . Only now do we spend effort on eigenvectors, because they fix the shape. For : gives ; check . For : gives . To sketch: draw both lines, put inward arrows on them, then draw curves that run parallel to the fast line far from the origin (running time backwards, the fast mode grows fastest) and bend to become tangent to the slow line at the origin (the fast mode has died). The sim’s preset draws exactly this.
Example 2: damping is a walk in the plane. The oscillator has companion matrix , so and , and the sim’s draggable point is literally the damping-stiffness plane (5.2 Constant Coefficients & the Characteristic Equation found the same cases from the characteristic roots). Fix and increase . At : , a center, (the undamped oscillator). At : , a stable spiral with (underdamped). At : , on the parabola, twice (critical damping, a degenerate node). At : , a stable node, (overdamped). Stability is and throughout, from and . Block E reproduces the four rows; to see them in the sim, drag the point to or , where the drag snaps to the parabola.
What to try. (1) Load the center preset: the readout gives a factor of per turn and closed ellipses; switch to center perturbed by +0.01 and the factor becomes while the ellipse opens into a spiral. (2) Back on the center, tick the Euler overlay: at the readout reports an amplitude gain of after about one period, at it reports , and RK4 with the same steps reports and (to six digits). (3) Pick the saddle and look at the dashed eigenlines: the curves hug them. (4) Tick the mirror box on a spiral and watch the spin reverse while , and the eigenvalues do not change.
Time reversal sends to , so and : the picture is mirrored across the -axis and every stable node becomes an unstable node, with the same curves traversed backwards. Mirroring the plane () keeps and and reverses the spin, which is why and not carries the direction. The sim’s checkbox does exactly that.
Moving across the plane: which pictures survive a perturbation?
Entries of a real matrix are never known exactly (measurement noise, rounding, model error), so ask which pictures survive a small change . and are continuous in the entries, so a point strictly inside a region stays inside: saddles, nodes and spirals are robust. Boundaries are different, because each is an equality:
- The center needs exactly. , so any perturbation with destroys it. Take the center and add to : now , , and the spiral factor is per turn (block B, where RK4 agrees to six digits, and the sim’s perturbed-center preset). A tiny outward spiral replaces closed orbits, and the sign of the perturbation chooses in or out. Honest caveat: some perturbations leave a center alone, namely those with ; putting in gives , , , still a center. But those are a thin slice of all perturbations.
- The parabola needs . Perturb by in : , , so . For the eigenvalues are and (a node), for they are (a spiral); block D and two sim presets. The sign of decides, so the degenerate node is a fragile hinge between two robust types.
- (an eigenvalue exactly zero) similarly splits into a saddle on one side and a node on the other.
This matters for modelling. A center in a model is a statement that the physics is exactly conservative; any unmodelled friction (or numerical error) moves the point off , and the long-term behaviour changes from “orbits forever” to “decays” or “blows up”. Which side you land on is decided by the sign of the error, as the next section shows for Euler.
Numerics: a classifier, and Euler on a center
Everything above is a function of the entries. The decision tree for the exercise is: means a line of equilibria; saddle; else look at : repeated (star if is a multiple of , otherwise degenerate); node; spiral, and a center when . The tolerance is a modelling decision: floating-point arithmetic almost never returns exactly, and you choose how close counts as “on the boundary”.
The test case the course cares about is the center. Apply explicit Euler to , (so ). One step is , and the quantity that the true flow conserves () is multiplied by
so the amplitude grows by every step, for every . In terms of the eigenvalue: the mode obeys and . Equivalently at , , : the Euler map expands area, though the true flow (, 7.5 The Matrix Exponential & Fundamental Matrices will show ) preserves it. Block C measures it over about one period :
| steps | Euler amplitude | formula | RK4 | |
|---|---|---|---|---|
| 0.1 | 31 | 1.8366 | 1.8366 | 0.99998629 |
| 0.05 | 63 | 1.3681 | 1.3681 | 0.99999956 |
| 0.05 | 628 (10 periods) | 22.7463 | 22.7463 | 0.99999564 |
The per-step factors are and . Halving roughly halves the logarithm of the gain per lap ( to ): the error is first order, consistent with Euler. After ten laps at the “orbit” has grown by a factor . RK4 at is within of a closed loop even after ten laps, because its amplification factor on the imaginary axis is rather than (from 4.3 Runge–Kutta Methods & Butcher Tableaux, the degree-4 Taylor polynomial of ). The sim’s Euler overlay draws the dotted outward spiral on the center and reports the same and for .
Computing it
Now build the classifier yourself. It must work from , and the entries only, and the tolerance is part of the specification.
Concept checks
Problems
Classify for . State the spin direction and the period.
Show worked solution
, . With and it is a center, . Spin: , so counterclockwise (the velocity at is , upward). Period . (These numbers are from the formula, not the sim.)
For , find , and classify for every real . Where are the structurally unstable values?
Show worked solution
, , so and the point travels along the horizontal line . For : stable node (e.g. : ). At : on the parabola, repeated (a degenerate node, since the matrix is not a multiple of ). For : stable spiral. At : center (). For : unstable spiral; at degenerate; for unstable node. The values are structurally unstable, the three crossings of the line with the parabola and the -axis. Stability: and means .
For show that the flow is incompressible and find the types for and .
Show worked solution
, so . , so . For : , : a saddle with . For : , : a center, . The sign of decides: strain beats rotation for a saddle, rotation beats strain for a center. (Numbers from the formula, not from the sim.)
Key Takeaways
- For the characteristic polynomial is , so , , .
- saddle; and node; spiral; center; degenerate or star node; a line of equilibria.
- Asymptotically stable exactly when and , with no eigenvalues needed.
- A spiral’s turn direction is the sign of (positive: counterclockwise), and its radius is multiplied by each turn ( for , measured ).
- The center and the parabola are structurally unstable: adding to of a center gives ( per turn), and in of a degenerate node gives a node or a spiral.
- Explicit Euler multiplies the amplitude of a center by per step (gain per lap at , over ten laps at ); RK4 at stays within of a closed loop after ten laps.
- In a 2D incompressible flow the trace of the Jacobian is the divergence, so and critical points are only saddles and centers (vortices); sources and sinks need .
- Next: complex and repeated eigenvalues solved exactly in 7.4 Complex & Repeated Eigenvalues, and the nonlinear case in 7.7 Nonlinear Systems & Linearization.