Visual walkthrough — Eigenvalues and eigenvectors — characteristic polynomial

Step 1 — What a matrix does to an arrow

WHAT. Start with the most basic object: an arrow (a vector) drawn from the origin. A matrix $A$ is a machine: feed it an arrow $\mathbf{x}$, and it spits out a new arrow, which we call $A\mathbf{x}$.

WHY. Before we can hunt for special arrows, we must see what happens to an ordinary one. Most arrows come out pointing somewhere new — they got turned.

PICTURE. In the figure, the cyan arrow is your input $\mathbf{x}$. The amber arrow is the output $A\mathbf{x}$. Notice the amber arrow points in a different direction — the machine rotated and lengthened it.

Term by term in $A\mathbf{x}$: the letter $A$ is the machine, the arrow $\mathbf{x}$ sits to its right as the thing being fed in, and writing them side-by-side means "apply $A$ to $\mathbf{x}$".

Step 2 — The magic directions where nothing turns

WHAT. Among all input arrows, a few special ones come out pointing along the same line they started on. The machine did not turn them — it only made them longer or shorter (or flipped them).

WHY. These are the arrows worth naming, because they reveal the "grain" of the matrix — the directions it treats simply. We call such an arrow an eigenvector, and the stretch factor an eigenvalue $\lambda$.

PICTURE. The cyan arrow $\mathbf{x}$ and the amber arrow $A\mathbf{x}$ now lie on the same dashed white line. The output is just the input scaled by a number $\lambda$ (here $\lambda =2$, so twice as long, same direction).

The equals sign is the whole demand: "the machine's output must land back on the input's line."

Step 3 — Why the zero arrow is forbidden

WHAT. Ask: what if we allowed $\mathbf{x}=0$, the arrow with no length?

WHY. Plug it in: $A0=0$ and $\lambda 0=0$ for every number $\lambda$. So the equation $A0=\lambda 0$ is true no matter what $\lambda$ is. That makes $\lambda$ meaningless — every number would "qualify". So we throw the zero arrow out.

PICTURE. The zero arrow is just a dot at the origin (drawn amber). No direction, no line to lie along — there is nothing for $\lambda$ to describe.

Step 4 — Move everything to one side

WHAT. Take $A\mathbf{x}=\lambda \mathbf{x}$ and subtract the right side across: $A\mathbf{x}-\lambda \mathbf{x}=0.$

WHY. We want to factor out the arrow $\mathbf{x}$, the way you factor $x$ out of $5x-3x$. To factor, all the $\mathbf{x}$ terms must sit together on one side, equal to zero.

PICTURE. Geometrically: if $A\mathbf{x}$ and $\lambda \mathbf{x}$ are the same arrow, then placing $\lambda \mathbf{x}$ tip-to-tail backwards against $A\mathbf{x}$ leaves you exactly at the origin — the closing white loop back to $0$.

• $A\mathbf{x}-\lambda \mathbf{x}$ ::: the output arrow minus the scaled input arrow.
• $=0$ ::: they cancel perfectly — the difference has no length.

Step 5 — The identity trick: turning $\lambda$ into a matrix

WHAT. We would love to write $(A-\lambda )\mathbf{x}$, but we cannot: $A$ is a grid of numbers (a matrix) and $\lambda$ is a single number. Subtracting a number from a grid is undefined — they live in different worlds. The repair: replace $\lambda$ with $\lambda I$, where $I$ is the identity matrix (the machine that changes nothing).

WHY. Because $I\mathbf{x}=\mathbf{x}$, multiplying by $\lambda I$ does the same job as multiplying by $\lambda$: $\lambda I\mathbf{x}=\lambda \mathbf{x}$. But now $\lambda I$ is a genuine grid, so $A-\lambda I$ is a legal matrix subtraction. See Determinants later; first we just need a real matrix.

PICTURE. The identity $I$ is drawn as $\mathbf{x}$ landing on itself (arrow unchanged). Scaling that by $\lambda$ gives the amber arrow $\lambda \mathbf{x}$ — identical to $\lambda I\mathbf{x}$.

$A\mathbf{x}-\lambda I\mathbf{x}=0⟹(A-\lambda I)\mathbf{x}=0.$

• $I$ ::: the identity matrix — $1$s down the diagonal, $0$s elsewhere; it leaves any arrow untouched.
• $\lambda I$ ::: a matrix that stretches every arrow by $\lambda$.
• $A-\lambda I$ ::: one honest matrix, formed by subtracting $\lambda$ from each diagonal entry of $A$.
• Factoring $\mathbf{x}$ out is now legal, since both terms are (matrix)$\times \mathbf{x}$.

Step 6 — When does a matrix crush an arrow to zero?

WHAT. Rename $M=A-\lambda I$. We need a nonzero arrow $\mathbf{x}$ with $M\mathbf{x}=0$. That means $M$ must collapse a real arrow down onto the origin.

WHY. If $M$ were reversible (invertible), we could undo it: multiply both sides by $M^{-1}$ to get $\mathbf{x}=M^{-1}0=0$ — forced back to the forbidden zero arrow! So $M$ must be the opposite of reversible: it must squash some direction to nothing. Such an $M$ has a nonzero null space. A collapsing matrix is called singular.

PICTURE. Left panel: an invertible $M$ maps the little square to a nonzero-area parallelogram — nothing collapses, only $0$ maps to $0$. Right panel: a singular $M$ flattens the whole square onto a line — a whole direction (cyan) is crushed onto the origin. That crushed direction is our eigenvector.

Step 7 — Singular means zero determinant

WHAT. How do we test whether $M=A-\lambda I$ collapses space, without drawing pictures? Use the determinant. The determinant $\det (M)$ measures the area (or volume) scaling of the machine.

WHY THIS TOOL. We need one number that screams "space got flattened". The determinant is exactly that number: if $M$ squashes the square onto a line, the enclosed area becomes $0$, so $\det (M)=0$. Nothing else so cleanly detects collapse — that's why the determinant, and not the trace or anything simpler, is the right instrument here.

PICTURE. The unit square (area $1$) maps to a parallelogram of area $|\det M|$. As $\lambda$ slides toward an eigenvalue, watch the parallelogram thin to a line — area $\to 0$.

$\det (A-\lambda I)=0.$

• $\det (\cdot )$ ::: the area/volume-scaling number of a matrix.
• $=0$ ::: the scaling collapsed — space was flattened — the matrix is singular.

Step 8 — Reading the equation as a curve

WHAT. Because the determinant of an $n\times n$ matrix is built from products of its entries, $p(\lambda )=\det (A-\lambda I)$ turns out to be a polynomial of degree $n$ in $\lambda$: the characteristic polynomial.

WHY. A polynomial is a curve we can graph, and its crossings of the horizontal axis are exactly the $\lambda$ that make $\det =0$ — the eigenvalues. So an abstract collapse condition becomes "find where a curve hits zero".

PICTURE. For $A=\left(\begin{array}{cc}2& 1\\ 1& 2\end{array}\right)$ we get $p(\lambda )={\lambda }^2-4\lambda +3$. The amber parabola crosses the axis at $\lambda =1$ and $\lambda =3$ — the two eigenvalues, marked with cyan dots.

• $p(\lambda )$ ::: the characteristic polynomial, height of the curve.
• roots of $p$ ::: the $\lambda$-values where the curve touches $0$ = the eigenvalues.

The sum of the roots is $1+3=4=\mathrm{tr}(A)$ and the product $1\cdot 3=3=\det (A)$ — the free sanity check from the trace and determinant. (See also Cayley–Hamilton Theorem: the matrix satisfies its own $p$.)

Step 9 — The degenerate cases (never left out)

WHAT. Three special situations the derivation must survive.

Case A — repeated root. $A=\left(\begin{array}{cc}3& 1\\ 0& 3\end{array}\right)$ gives $p(\lambda )=(3-\lambda {)}^2$: the curve touches the axis at $\lambda =3$ without crossing (a double root). Algebraic multiplicity $2$, but only one eigenvector direction.

Case B — no real roots. The rotation $A=\left(\begin{array}{cc}0& -1\\ 1& 0\end{array}\right)$ gives $p(\lambda )={\lambda }^2+1$: the parabola floats above the axis, never touching. No real eigenvalue — the roots are $\pm i$, living in the complex world. A real matrix that turns every arrow has no real eigen-direction, exactly as the picture warns.

Case C — the identity itself. $A=I$ gives $A\mathbf{x}=\mathbf{x}$ for every arrow: $\lambda =1$ works in all directions. $p(\lambda )=(1-\lambda {)}^2$, and every nonzero arrow is an eigenvector.

PICTURE. Three parabolas on shared axes: crossing twice (distinct real), touching once (repeated), floating clear (complex).

The one-picture summary

Every arrow shows one step: the eigen-equation $\to$ move to one side $\to$ insert $I$ $\to$ demand collapse $\to$ determinant zero $\to$ read off the curve's roots.

Flashcards