Chapter 9 — Key Takeaways

The one idea

The inverse matrix $A^{-1}$ is the transformation that undoes $A$, and it exists exactly when $A$ loses no dimension. If $A$ rotates, stretches, shears, or reflects space, then $A^{-1}$ retraces that motion in reverse, so that $A^{-1}A = AA^{-1} = I$ (the do-nothing transformation). A matrix has an inverse precisely when its transformation is bijective — never crushing two inputs to the same output, never failing to reach a target — which for a square matrix is the single geometric fact behind a long list of equivalent algebraic conditions. Singular (non-invertible) matrices are the ones that collapse space onto a lower-dimensional shadow, destroying the information needed to reverse them.

The big ideas, in order

1. The inverse is the undo. $A^{-1}A = AA^{-1} = I$. Geometry first: reverse the stretch, un-rotate, un-shear. A reflection is its own inverse; a rotation's inverse is the opposite rotation (and equals its transpose); a scaling inverts by reciprocals; a shear by negating its slide.
2. Invertible $\iff$ bijective $\iff$ no dimension lost. A reversible transformation must be one-to-one (no two inputs collide) and onto (every output reached). For square matrices these coincide, so you check only one.
3. Singular = collapse = irreversible. A singular matrix flattens $\mathbb{R}^n$ onto a line/plane/point; infinitely many inputs pile onto each output, so no undo exists. Tells: a zero column, a repeated column, a column that's a multiple of another, a nonzero null space, $\det = 0$.
4. Compute $A^{-1}$ by Gauss-Jordan on $[A\,|\,I]$. Row-reduce until the left block is $I$; the right block becomes $A^{-1}$. A missing pivot (zero row) means singular — no inverse. For $2\times 2$: $\begin{bmatrix}a&b\\c&d\end{bmatrix}^{-1} = \frac{1}{ad-bc}\begin{bmatrix}d&-b\\-c&a\end{bmatrix}$, valid iff $ad - bc \neq 0$.
5. The Invertible Matrix Theorem. For square $A$, these are all equivalent: invertible; bijective; full rank $n$; pivot in every row/column; independent columns; $\det \neq 0$; trivial null space; unique solution to $A\mathbf{x}=\mathbf{b}$ for every $\mathbf{b}$. One geometric idea, eight faces — pick the cheapest to check.
6. Properties, with the order-reversal trap. $(A^{-1})^{-1} = A$; $(AB)^{-1} = B^{-1}A^{-1}$ (reverse the order — socks then shoes, shoes off first); $(A^{\mathsf{T}})^{-1} = (A^{-1})^{\mathsf{T}}$; $(A^k)^{-1} = A^{-k}$; $\det(A^{-1}) = 1/\det(A)$.
7. Rarely invert to solve. Write $\mathbf{x} = A^{-1}\mathbf{b}$ on paper, but in code use np.linalg.solve(A, b) — it's faster (about a third the work) and more accurate. Reserve inv for when you need the inverse as an object (multipliers, precision matrices, reused color/coordinate transforms).

The invertibility checklist at a glance

Skills you gained

• Describe $A^{-1}$ geometrically as the reverse motion, and write inverses of the standard transformations by inspection.
• Decide invertibility by whichever face of the theorem is cheapest (spot a dependent column, compute a $2\times 2$ determinant, count pivots).
• Compute an inverse by Gauss-Jordan on $[A\,|\,I]$, by hand for $2\times 2$/$3\times 3$ and in numpy for any size; recognize singularity as a missing pivot.
• Apply and justify the inverse properties, especially the order-reversal $(AB)^{-1} = B^{-1}A^{-1}$.
• Choose solve over inv to solve systems, and know the few cases where the inverse-as-object is worth computing.
• Implement inverse(A) from scratch (Gauss-Jordan, reusing Chapter 4's row reduction), verified against numpy.

Terms to know

inverse matrix, invertible (nonsingular) matrix, singular (non-invertible) matrix, identity matrix, Gauss-Jordan elimination, augmented matrix $[A\,|\,I]$, bijective (injective + surjective), full rank, pivot, determinant, the Invertible Matrix Theorem, two-sided / left / right inverse, condition number (named, deferred to Chapter 38), Leontief inverse, modular inverse.

How this connects to the recurring themes

• Theme 1 (transformations are the point). The inverse is the reverse transformation; this chapter adds a past tense to Chapter 7's verb. Invertibility is a property of the transformation, not the coordinate grid.
• Theme 2 (geometry = algebra). The Invertible Matrix Theorem is this theme: "loses no dimension," "$\det \neq 0$," "full rank," "trivial null space," and "independent columns" are one geometric fact in many algebraic dialects.
• Theme 3 (computation validates theory). Every inverse matched numpy; your toolkit gained a from-scratch inverse; and §9.8's "solve, don't invert" is the theme made into a daily coding habit.
• Theme 5 (the four fundamental subspaces). The null space surfaced here as "the vectors a singular matrix kills" — a trivial null space is exactly invertibility, foreshadowing Part III's structural picture.

Toolkit contribution

toolkit/inverse.py — inverse(A) via Gauss-Jordan on $[A\,|\,I]$, reusing row_reduce from Chapter 4's toolkit/linear_systems.py, returning $A^{-1}$ or raising ValueError on a singular matrix; verified against np.linalg.inv. Chapter 10 reuses the same elimination yet again to build the LU factorization — the efficient way to solve systems we championed in §9.8.

Forward references

• Chapter 10 — LU and PLU decomposition: the efficient way to solve systems (and the real engine inside np.linalg.inv), built from this chapter's elimination — cashing in §9.8's "don't invert, factor."
• Chapter 11 — The determinant as signed volume scaling; the rigorous proof that $\det(A) = 0 \iff$ singular, and $\det(A^{-1}) = 1/\det(A)$.
• Chapter 13–14 — The null space and the four fundamental subspaces; "trivial null space" placed inside the structural theory, with Rank–Nullity.
• Chapter 16 — Change of basis, where $P^{-1}AP$ uses the inverse to re-express a transformation in new coordinates.
• Chapter 19 — The pseudoinverse $(A^{\mathsf{T}}A)^{-1}A^{\mathsf{T}}$: the one-sided inverse that solves least squares for non-square matrices.
• Chapter 21 — Orthogonal matrices, where $A^{-1} = A^{\mathsf{T}}$ exactly (the rotation surprise from §9.5, explained).
• Chapter 25 — Diagonalization $A = PDP^{-1}$ and matrix powers $A^k = PD^kP^{-1}$, where the inverse appears symbolically and powers invert by negating exponents.
• Chapter 38 — The condition number: the right measure of "near-singular," replacing the misleading "small determinant."