The Invertible Matrix Theorem Explained: 50 Equivalent Conditions That Connect Linear Algebra

Part I — The Theorem, the Count, and the Two Worlds

Why this may be the first theorem to learn

A typical linear algebra course introduces ideas one at a time: elimination, vector spaces, transformations, determinants, eigenvalues, orthogonality, and perhaps singular values. That sequence is necessary, but it can make the subject feel fragmented. The Invertible Matrix Theorem supplies the missing architecture.

Every major condition answers one of two deeper questions:

1. Does \(A\) destroy information? This is the kernel, independence, injectivity, smallest-singular-value side.
2. Can \(A\) reach every target? This is the column space, spanning, consistency, surjectivity side.

For a square map \(\mathbb R^n\to\mathbb R^n\), these questions have the same answer. If no direction is lost, all \(n\) output directions remain available. If every output is reachable, there is no room for two inputs to produce the same output. That equality of dimensions is the engine behind the theorem.

Are there 9, 23, 24, or 50 conditions?

All of those counts can be defensible because authors group statements differently. Some textbooks give a compact core of 8–12 conditions. MathWorld’s Invertible Matrix Theorem presents 23 numbered conditions equivalent to the ambient statement that \(A\) is invertible. Counting that ambient statement separately produces 24 propositions.

This lesson uses a declared teaching rule: it counts 50 pedagogically useful and independently searchable formulations. Some are close vocabulary restatements—such as \(\operatorname{Nul}(A)=\{0\}\) and \(\operatorname{nullity}(A)=0\)—because students encounter and search for both. There is no universal largest number: infinitely many equivalent statements can be manufactured.

The all-true-or-all-false principle

For

every condition in the atlas is true. The matrix has two pivots, trivial null space, rank two, a two-sided inverse, nonzero eigenvalues, positive singular values, and a transformation that is one-to-one and onto.

For

every condition is false. Its second column is twice its first; \(B[-2,1]^T=0\); rank is one; zero is an eigenvalue and singular value; and the map collapses the plane onto a line.

This binary split is the practical power of the theorem. You do not need to verify fifty statements. Find the cheapest decisive test available, then invoke equivalence.

Part II — The Square-Matrix Hypothesis

The standard theorem assumes

The word square is not a technicality. It makes row conditions and column conditions meet in the middle.

What changes for rectangular matrices?

Consider a \(3\times2\) matrix \(C\). It can have a pivot in every column and define a one-to-one map \(\mathbb R^2\to\mathbb R^3\), yet it cannot be onto \(\mathbb R^3\) because two input directions cannot span three output dimensions.

Now consider a \(2\times3\) matrix \(D\). It can have a pivot in every row and define an onto map \(\mathbb R^3\to\mathbb R^2\), yet it cannot be one-to-one because rank–nullity forces a nontrivial kernel.

The Moore–Penrose pseudoinverse is useful for rectangular and rank-deficient matrices, but it is not an ordinary two-sided inverse and does not make the square IMT statements true.

What changes over the complex numbers?

The algebraic core works over any field with suitable language. Over \(\mathbb C\), replace the transpose by the conjugate transpose in inner-product statements:

Then \(x^*A^*Ax=\|Ax\|_2^2\), and the positive-definite Gram-matrix formulation remains correct.

Part III — The Complete 50-Condition Atlas

Every statement in the following atlas is equivalent for \(A\in\mathbb R^{n\times n}\). The “why it matches” column is intentionally short: use it for fast review, then visit the detailed family sections for proofs and examples.

How to read the atlas without memorizing fifty lines

Compress it into six hubs:

1. Pivots: \(n\) pivots and \(\operatorname{rref}(A)=I_n\).
2. Kernel and image: \(\operatorname{Nul}(A)=\{0\}\) and \(\operatorname{Col}(A)=\mathbb R^n\).
3. Systems and maps: exactly one solution for every \(b\); one-to-one and onto.
4. Determinant and spectrum: \(\det(A)\ne0\); zero is not an eigenvalue.
5. SVD and Gram matrix: \(\sigma_{min}>0\); \(A^TA\) is positive definite.
6. Geometry: no independent direction or positive-dimensional volume is collapsed.

If you can prove the bridges among those six hubs, every numbered formulation becomes a definition, a vocabulary translation, or a short corollary.

Part IV — Inverse, Row Reduction, Pivots, and Elementary Matrices

Conditions 1–6: the computational engine

An inverse \(A^{-1}\) satisfies

Row operations are reversible. If elimination carries \(A\) to \(I_n\), then the product of the corresponding elementary matrices \(E_k\cdots E_1\) satisfies

Therefore \(A^{-1}=E_k\cdots E_1\). Conversely, if elimination stops with a missing pivot, the reduced matrix cannot be \(I_n\), so \(A\) cannot have a two-sided inverse.

For an \(n\times n\) matrix, the following are the same event:

The augmented-matrix inverse algorithm

Row reducing

performs the same reversible operations on both blocks. If the left block reaches \(I_n\), the right block becomes \(A^{-1}\):

If the left side cannot reach the identity, the inverse does not exist. This algorithm simultaneously tests invertibility and computes the inverse.

Conditions 29–31: left inverse, right inverse, and elementary factorization

For square matrices, \(BA=I_n\) implies \(A\) is injective: if \(Ax=0\), then \(x=BAx=0\). Injectivity implies invertibility, so \(B=A^{-1}\) and automatically \(AB=I_n\). Similarly, a right inverse \(AC=I_n\) forces surjectivity and hence invertibility.

Each row operation equals multiplication by an elementary matrix. Thus \(A\) is invertible exactly when it is a product of elementary matrices. This is more than a construction trick: it says an invertible matrix is built entirely from reversible operations—row swaps, nonzero scalings, and row replacements.

Part V — Null Space, Independence, Spanning, Bases, Rank, and Nullity

Conditions 7–10: no free variable means no hidden direction

Write the columns of \(A\) as \(a_1,\ldots,a_n\). The homogeneous equation

is the vector equation

The columns are linearly independent exactly when the only coefficients producing zero are \(x_1=\cdots=x_n=0\). Therefore

Elimination explains the same chain mechanically. A pivot in every column means no free variable. No free variable means only the zero solution in the homogeneous system.

Conditions 11–15: reaching every direction

The equation \(Ax=b\) asks whether \(b\) is a linear combination of the columns. Therefore \(Ax=b\) is consistent for every \(b\in\mathbb R^n\) exactly when

That is the same as saying the columns span \(\mathbb R^n\). Because there are exactly \(n\) columns in an \(n\)-dimensional space, an independent list of \(n\) vectors automatically spans, and a spanning list of \(n\) vectors is automatically independent. Hence the columns form a basis.

Rank packages the image dimension into one number:

Rank–nullity packages the input-side accounting:

Thus rank \(n\) and nullity zero are two sides of the same dimension equation.

Existence and uniqueness are different ideas

Students often merge them, but the theorem keeps them distinct before joining them:

The difference of any two solutions lies in the null space. If \(Ax_1=b\) and \(Ax_2=b\), then

So uniqueness for every consistent system is exactly the statement that the null space is trivial.

Part VI — Systems and Linear Transformations

Conditions 23–28: translate equations into maps

The matrix equation \(Ax=b\) and the transformation equation \(T(x)=b\) are the same statement. Consequently:

• \(Ax=b\) is consistent for every \(b\) iff \(T\) is onto.
• \(Ax=b\) has at most one solution for every \(b\) iff \(T\) is one-to-one.
• \(Ax=b\) has exactly one solution for every \(b\) iff \(T\) is bijective.
• \(T\) is bijective iff an inverse transformation \(T^{-1}\) exists.

When \(T(x)=Ax\) is invertible, its inverse is represented by \(A^{-1}\):

Why one-to-one and onto become equivalent

For a general function, injectivity and surjectivity are different. For a rectangular linear map, they are still different. For a linear map from an \(n\)-dimensional space to another \(n\)-dimensional space, rank–nullity forces them together.

If \(T\) is one-to-one, then \(dimker T=0\), so

The range is an \(n\)-dimensional subspace of \(\mathbb R^n\), so it must be the entire codomain: \(T\) is onto.

If \(T\) is onto, then the range already has dimension \(n\). Rank–nullity leaves zero dimensions for the kernel, so \(T\) is one-to-one.

The coordinate-change interpretation

A basis matrix \(P=[v_1\ \cdots\ v_n]\) is invertible exactly when its columns form a basis. Then

Invertibility is what makes coordinates unique and reversible. If the columns were dependent, one vector could have multiple coordinate lists; if they did not span, some vectors would have no coordinates in that list at all.

Part VII — Determinants, Geometry, Eigenvalues, and Polynomials

Condition 34: the determinant test

The determinant converts the entire theorem into one scalar test:

Algebraically, a zero determinant signals dependent rows or columns and a missing pivot. Geometrically, \(|\det(A)|\) is the factor by which \(A\) scales \(n\)-dimensional volume. A nonzero factor means an \(n\)-dimensional region remains \(n\)-dimensional; zero means it collapses into a lower-dimensional set.

The sign of \(\det(A)\) records orientation. A positive determinant preserves orientation, and a negative determinant reverses it. Nonzero alone guarantees invertibility; it does not guarantee orientation preservation.

For a deeper geometric treatment, see Determinants Explained: Algebra, Geometry, and Differential Equations.

Condition 35: zero is not an eigenvalue

By definition, \(\lambda\) is an eigenvalue if there is a nonzero vector \(v\) with

Setting \(\lambda=0\) gives \(Av=0\) for some \(v\ne0\). Therefore

So an invertible matrix may have positive, negative, real, or complex eigenvalues, but none can equal zero. Invertibility does not imply diagonalizability; a matrix can be invertible and defective.

For the complete spectral foundation, see Eigenvalues and Eigenvectors Explained.

Conditions 36–38: characteristic and minimal polynomials

Using the convention

we obtain

Thus \(p_A(0)\ne0\) exactly when \(\det(A)\ne0\). The sign factor matters in odd dimensions, even though the zero-versus-nonzero test is unchanged.

The minimal polynomial \(m_A\) also detects invertibility. If \(m_A(0)=0\), then \(m_A(t)=tq(t)\), so \(Aq(A)=0\); the minimality structure forces zero into the spectrum. If \(m_A(0)\ne0\), write

Since \(m_A(A)=0\),

Therefore \(A^{-1}\) is a polynomial in \(A\). Cayley–Hamilton gives a similar construction from the characteristic polynomial.

Part VIII — Rows, Transpose, and the Four Fundamental Subspaces

Conditions 16–22: row–column duality

The columns of \(A^T\) are the rows of \(A\). Since

full column rank for \(A\) is the same as full column rank for \(A^T\). That immediately connects invertibility of \(A\), invertibility of \(A^T\), independent rows, spanning rows, a row basis, and zero transpose nullity.

Also,

because transposing \(AA^{-1}=I\) reverses the product order.

The four fundamental subspaces

For \(A\in\mathbb R^{m\times n}\), the four fundamental subspaces are

• \(\operatorname{Col}(A)\subseteq\mathbb R^m\),
• \(\operatorname{Nul}(A^T)\subseteq\mathbb R^m\),
• \(\operatorname{Row}(A)=\operatorname{Col}(A^T)\subseteq\mathbb R^n\),
• \(\operatorname{Nul}(A)\subseteq\mathbb R^n\).

They pair through orthogonal complements:

For an invertible \(n\times n\) matrix, both null spaces are zero and both row and column spaces are all of \(\mathbb R^n\). Conditions 32 and 33 are direct restatements of those orthogonal-complement identities.

Why this matters for solvability

For any matrix—not necessarily square—the system \(Ax=b\) is consistent exactly when \(b\in\operatorname{Col}(A)\). Equivalently, \(b\) must be orthogonal to every vector in \(\operatorname{Nul}(A^T)\). This is the finite-dimensional Fredholm alternative:

When \(A\) is invertible, \(\operatorname{Nul}(A^T)=\{0\}\), so there is no compatibility obstruction and every \(b\) is reachable.

Part IX — SVD, Gram Matrices, Lower Bounds, and Conditioning

Conditions 39–40: singular values

Every real matrix has a singular value decomposition

where (U) and (V) are orthogonal and the diagonal entries of (Sigma) are the singular values

Orthogonal factors only rotate or reflect. The singular values measure stretches along perpendicular directions. A zero singular value means one direction is flattened completely. Thus

Conditions 41–44: the Gram matrix

The Gram matrix is \(A^TA\). Its central identity is

Therefore \(x^TA^TAx>0\) for every \(x\ne0\) exactly when \(Ax\ne0\) for every \(x\ne0\). That is injectivity, which is invertibility for a square matrix. Hence

The eigenvalues of \(A^TA\) are \(\sigma_i^2\). Consequently the Gram determinant satisfies

It is positive exactly when no singular value is zero.

Condition 45: bounded below

The correct equivalent statement is:

This says that \(A\) cannot shrink any nonzero input arbitrarily close to zero relative to its size. The best possible constant is \(c=\sigma_{\min}(A)\).

Be careful: the inequality \(\|Ax\|_2\ge\sigma_{\min}(A)\|x\|_2\) is true for every matrix. It becomes an invertibility test only when the lower-bound constant is required to be strictly positive.

Condition 46: condition number

The spectral condition number is

It is finite exactly when \(A\) is invertible. But “finite” can still mean enormous. A large condition number signals that relative errors in data or rounding can be strongly amplified.

Critical distinction: nonzero \(\sigma_{\min}\) decides exact invertibility. The ratio \(\sigma_{\max}/\sigma_{\min}\) measures conditioning. The scale-dependent statement “\(\sigma_{\min}\) is small” does not by itself mean ill-conditioned; \(10^{-6}I\) has a small smallest singular value but \(\kappa_2=1\).

In spectral norm, the distance from \(A\) to the nearest singular matrix equals \(\sigma_{\min}(A)\). This explains why the smallest singular value is a safety margin while the condition number is a relative sensitivity measure.

Part X — Complete Worked Examples

Invertible example: activate the whole theorem

Let

Its determinant is

That single calculation is enough to invoke all fifty conditions. To see the connections rather than merely cite them, row reduce:

There are two pivots and no free variables. Therefore the columns are independent. Because there are two independent columns in \(\mathbb R^2\), they form a basis and span the plane. The null space is \(\{0\}\), rank is two, nullity is zero, and the associated map is one-to-one and onto.

The inverse is

For \(b=[5,3]^T\),

and \(Ax=[5,3]^T\) verifies the answer.

Its characteristic polynomial is

with eigenvalues

Both are nonzero. Because this particular \(A\) is symmetric positive definite, its singular values equal its positive eigenvalues: approximately \(2.618\) and \(0.382\). Thus

Also,

and the unit square maps to a parallelogram of area \(|\det A|=1\).

Singular example: collapse the whole theorem

Now let

Its determinant is zero and its second column is twice its first. Row reduction gives

There is one pivot and one free variable. Indeed,

So the null space is nontrivial, the columns are dependent, rank is one, nullity is one, and the map is not one-to-one. Its column space is only a line, so the map is not onto \(\mathbb R^2\). Some right-hand sides have no solution, while consistent systems have infinitely many solutions.

Zero is an eigenvalue and singular value. The unit square collapses to a line segment with area zero. No inverse exists. Again, one decisive failure reveals the entire world.

Side-by-side ledger

Part XI — The Proof Network

Proof spine 1: elimination

Reversible row operations provide both directions and give the elementary-matrix factorization.

Proof spine 2: pivots to kernel and image

The same \(n\) pivots include a pivot in every row:

Rank–nullity and the basis theorem attach rank, nullity, and basis language.

Proof spine 3: systems and transformations

while

The square equal-dimension hypothesis turns one-to-one into onto and combines existence with uniqueness.

Proof spine 4: determinant and spectrum

Elimination or the adjugate identity establishes

The eigenvalue definition establishes

Evaluating \(p_A(t)=\det(tI-A)\) at zero and using Cayley–Hamilton attaches the polynomial conditions.

Proof spine 5: transpose and subspaces

Rank is unchanged by transpose, so full row rank and full column rank coincide. The identities

attach the orthogonality conditions.

Proof spine 6: SVD and Gram matrices

From

positive definiteness of \(A^TA\) is exactly triviality of the kernel. The eigenvalues of \(A^TA\) are \(\sigma_i^2\), so positive definiteness, positive singular values, a positive lower bound, and finite \(\kappa_2\) meet at \(\sigma_{\min}>0\).

Proof spine 7: geometry and dimension

The transformed unit cube has volume \(|\det A|\), so positive volume equals nonzero determinant. An injective linear map preserves independent sets and the dimension of every subspace. Conversely, if every one-dimensional subspace keeps dimension one, no nonzero vector can map to zero, so the kernel is trivial.

A short circular proof worth memorizing

If your instructor requests a compact proof of the classical core, use this loop:

Then attach the inverse, determinant, basis, rank, spectrum, transpose, and SVD families through their defining bridge identities.

Part XII — Fastest Tests, Applications, Mistakes, and Mastery

Which invertibility test should you use?

In numerical software, avoid deciding near-singularity from exact equality with a floating-point determinant. Rank-revealing factorizations and singular values are safer because they expose scale and sensitivity.

Applications across linear algebra and beyond

Coordinate changes. A change-of-basis matrix must be invertible so coordinates can move both directions without ambiguity.

Differential equations. For \(x^{\prime}=Ax\), an invertible \(A\) means zero is not an eigenvalue, so the origin has no center direction caused by a zero eigenvalue. Stability still depends on the real parts of all eigenvalues, not invertibility alone. See Phase Portraits, Eigenvalues, and Stability.

Least squares. If a rectangular design matrix \(X\) has independent columns, \(X^TX\) is invertible and the normal equations have the unique solution

Notice the distinction: \(X\) itself need not be square; invertibility is asserted for the square Gram matrix \(X^TX\).

Eigenvector matrices. A matrix \(P\) of eigenvectors is invertible exactly when those eigenvectors form a basis. Then \(A=PDP^{-1}\). An invertible \(A\) need not be diagonalizable; these are separate properties.

Numerical linear algebra. Solving \(Ax=b\) depends not only on whether \(A^{-1}\) exists in exact arithmetic but also on conditioning. A huge \(\kappa(A)\) can make computed solutions sensitive to small perturbations.

Information and data transformations. An invertible linear layer has no collisions: each output comes from exactly one input. Rank deficiency means dimensional information is destroyed.

Advanced equivalent corollaries

The following are useful extensions rather than part of the headline count:

• \(A^k\) is invertible for some positive integer \(k\), equivalently for every positive integer \(k\).
• For every \(D\in M_n(\mathbb R)\), \(AX=D\) has exactly one solution \(X\).
• For every \(D\in M_n(\mathbb R)\), \(XA=D\) has exactly one solution \(X\).
• Left multiplication \(X\mapsto AX\) and right multiplication \(X\mapsto XA\) are invertible operators on \(M_n(\mathbb R)\).
• Matrix cancellation by \(A\) is valid on both sides.
• For \(n\ge2\), \(\operatorname{adj}(A)\) is invertible.
• \(A\) lies a positive spectral-norm distance from the singular matrices, equal to \(\sigma_{\min}(A)\).
• The reciprocal condition number is positive.

These examples also prove why no final universal count exists.

Common mistakes and exam traps

1. Forgetting the square hypothesis. One-to-one does not imply onto for a tall matrix, and onto does not imply one-to-one for a wide matrix.
2. Saying “the determinant exists, so the matrix is invertible.” Every square matrix has a determinant; it must be nonzero.
3. Confusing “consistent” with “unique.” Existence is controlled by the column space; uniqueness is controlled by the null space.
4. Checking only rows or only columns without using squareness. Pivot-every-row and pivot-every-column coincide only in the square full-rank case.
5. Believing nonzero eigenvalues imply diagonalizability. They imply invertibility, not a complete eigenbasis.
6. Believing invertible means numerically safe. A matrix can have a positive but tiny relative safety margin and a huge condition number.
7. Using \(A^T\) instead of \(A^*\) over \(\mathbb C\). Inner-product and Gram statements require conjugation.
8. Computing \(A^{-1}\) when only a solution is needed. Solve \(Ax=b\) by factorization; explicitly forming the inverse is usually slower and less stable.
9. Treating a one-sided inverse of a rectangular matrix as two-sided. Dimensions and rank determine which side can exist.
10. Dropping the sign convention in the characteristic polynomial. For \(p_A(t)=\det(tI-A)\), \(p_A(0)=(-1)^n\det(A)\).

The Woody mastery method

Do not memorize fifty isolated sentences. Practice translation from one dialect to another:

1. Start with one condition, such as “\(Ax=0\) has only the trivial solution.”
2. Translate it into kernel language: \(\operatorname{Nul}(A)=\{0\}\).
3. Translate kernel into columns: the columns are independent.
4. Use squareness: \(n\) independent columns in \(\mathbb R^n\) form a basis and span.
5. Translate span into systems and maps: every \(b\) is reachable, so \(T\) is onto.
6. Return to computation: full span means a pivot in every row; square size gives \(n\) pivots and RREF \(I_n\).
7. Attach determinant, eigenvalue, SVD, and geometry tests.

Your goal is to rebuild the network from any starting point. If you can explain three different routes from a given condition to invertibility, you understand the theorem rather than merely recognizing its wording.

Part XIII — Linear Algebra Reference Dictionary

This dictionary makes the lesson self-contained for a student meeting some of the vocabulary for the first time. Each entry gives a definition, its place in the theorem, and the fastest bridge to invertibility.

Inverse matrix

An inverse of a square matrix \(A\) is a matrix \(A^{-1}\) satisfying

Multiplying by \(A\) performs a reversible linear transformation; multiplying by \(A^{-1}\) reverses it. The inverse is unique. If \(B\) and \(C\) were both inverses, then

Useful rules include

The reverse order in the product rule matters because matrix multiplication is generally not commutative.

Row echelon form, reduced row-echelon form, and pivots

A matrix is in row echelon form when nonzero rows appear above zero rows, each leading entry moves right as you move down, and entries below each leading entry are zero. Reduced row-echelon form adds two conditions: each leading entry is \(1\), and it is the only nonzero entry in its column.

A pivot position is the location of a leading entry in an echelon form. Pivots reveal rank and identify basic variables. Columns without pivots correspond to free variables in \(Ax=0\). For a square matrix, no free variables means a pivot in every column; \(n\) pivots then also means a pivot in every row.

Homogeneous and nonhomogeneous systems

A system \(Ax=0\) is homogeneous and is always consistent because \(x=0\) is a solution. Its decisive question is uniqueness: are there nonzero solutions? Those nonzero solutions are precisely the information-losing directions of \(A\).

A system \(Ax=b\) is nonhomogeneous when \(b\ne0\). Its decisive questions are existence and uniqueness. Existence asks whether \(b\) lies in the column space. Uniqueness asks whether the homogeneous system has only the zero solution.

If \(x_p\) is one particular solution, every solution has the form

This formula explains the three standard possibilities: no solution, exactly one solution, or infinitely many solutions. For an invertible coefficient matrix, only the exactly-one case occurs, for every \(b\).

Null space or kernel

The null space is

It records every direction erased by the transformation. A trivial null space means no nonzero vector is erased, so the map is one-to-one. A nontrivial null space gives a collision: if \(Az=0\) with \(z\ne0\), then \(A(x+z)=Ax\).

Column space or range

The column space is

It is the set of all reachable outputs and also the span of the columns. The system \(Ax=b\) is solvable precisely when \(b\in\operatorname{Col}(A)\). For a square invertible matrix, the column space is all of \(\mathbb R^n\).

Linear independence

Vectors \(v_1,\ldots,v_k\) are linearly independent if

forces \(c_1=\cdots=c_k=0\). Applied to the columns of \(A\), that equation is \(Ax=0\). Column independence and a trivial null space are therefore exactly the same statement written in two dialects.

Span

The span of a list of vectors is the set of all their linear combinations. The columns of \(A\) span \(\mathbb R^n\) when every \(b\in\mathbb R^n\) can be expressed as

Thus spanning columns, universal consistency of \(Ax=b\), full column space, and onto behavior are the same fact.

Basis and coordinates

A basis is a linearly independent spanning set. Independence gives unique coordinates; spanning guarantees that coordinates exist. A square matrix is invertible precisely when its columns form a basis of \(\mathbb R^n\). The inverse converts standard coordinates back into coordinates relative to that basis.

The basis theorem is a key finite-dimensional shortcut: every independent list of \(n\) vectors in \(\mathbb R^n\) is automatically a basis, and every spanning list of \(n\) vectors is automatically a basis.

Rank and nullity

Rank is the dimension of the image:

Nullity is the dimension of the kernel:

For a matrix with \(n\) columns, rank–nullity says

Invertibility is the extreme full-information case: rank \(n\) and nullity \(0\).

One-to-one, onto, and bijective

A map is one-to-one if different inputs always produce different outputs. For a linear map, this is equivalent to a trivial kernel.

A map is onto if every target in the codomain is reached. For a matrix map, this is equivalent to the column space equaling the codomain.

A map is bijective if it is both one-to-one and onto. Bijectivity is the function-level definition of invertibility. The special square finite-dimensional fact is that either one-to-one or onto alone forces the other.

Determinant

The determinant is a signed volume-scaling factor. It is multilinear in the rows or columns, changes sign when two rows are swapped, and vanishes when rows or columns are dependent. Its multiplication rule,

immediately shows that if \(A\) is invertible, then

so \(\det(A)\ne0\) and \(\det(A^{-1})=1/\det(A)\).

Eigenvalue and spectrum

An eigenvalue \(\lambda\) is a scalar for which \(Av=\lambda v\) for some nonzero eigenvector \(v\). The spectrum is the set of eigenvalues, counted with context-appropriate multiplicity. The only eigenvalue that decides invertibility by itself is zero:

Other spectral properties answer different questions. For example, diagonalizability asks whether there are enough independent eigenvectors, while stability of \(x^{\prime}=Ax\) depends on the real parts of eigenvalues.

Singular value

The singular values are the nonnegative square roots of the eigenvalues of \(A^TA\). Unlike eigenvalues, they are always real and nonnegative. They measure the principal stretch factors of the transformation. Rank equals the number of positive singular values, and invertibility requires all \(n\) to be positive.

Gram matrix and positive definiteness

The Gram matrix \(A^TA\) stores all column inner products. Its \((i,j)\) entry is \(a_i^Ta_j\). It is always symmetric and positive semidefinite because

It is positive definite precisely when equality can occur only at \(x=0\), which is precisely when the columns of \(A\) are independent.

Condition number

The condition number estimates relative sensitivity. In the Euclidean norm,

An orthogonal matrix has condition number \(1\), the best possible value. A singular matrix has infinite condition number under the standard convention. An invertible matrix with a very large condition number is mathematically reversible but numerically delicate.

Part XIV — Implication Recipes and Solved Practice

These examples demonstrate how to use the theorem efficiently. Each solution names the shortest bridge and then unlocks the requested conclusion.

Recipe 1: from a homogeneous system to every right-hand side

Given: \(A\in\mathbb R^{n\times n}\) and \(Ax=0\) has only \(x=0\).
Prove: \(Ax=b\) has exactly one solution for every \(b\in\mathbb R^n\).

The given condition says \(\operatorname{Nul}(A)=\{0\}\), so \(A\) is one-to-one. Because the map is square and finite-dimensional, one-to-one implies onto. Onto gives existence for every \(b\), and the trivial kernel gives uniqueness. Therefore every system has exactly one solution.

Recipe 2: from spanning rows to a nonzero determinant

Given: the rows of \(A\in\mathbb R^{n\times n}\) span \(\mathbb R^n\).
Prove: \(\det(A)\ne0\).

Spanning rows give \(\operatorname{Row}(A)=\mathbb R^n\), so row rank is \(n\). Row rank equals column rank, hence \(\operatorname{rank}(A)=n\). Full rank is equivalent to invertibility, and an invertible matrix has nonzero determinant.

Recipe 3: from a left inverse to unique solvability

Given: \(BA=I_n\).
Prove: \(Ax=b\) has exactly one solution for every \(b\).

If \(Ax=0\), multiply by \(B\):

Thus \(A\) has trivial kernel and is injective. Since \(A\) is square, it is invertible. In fact \(B=A^{-1}\), so the unique solution is \(x=Bb\).

Recipe 4: from a Gram matrix to independent columns

Given: \(A^TA\) is positive definite.
Prove: the columns of \(A\) are linearly independent.

If \(Ax=0\), then

Positive definiteness allows equality only for \(x=0\). Therefore the null space is trivial, which is exactly column independence.

Recipe 5: from a singular value to volume

Given: \(\sigma_{\min}(A)>0\).
Prove: the image of the unit cube has positive volume.

Positive smallest singular value means every singular value is positive, so their product is positive. Since

the determinant is nonzero. The transformed unit cube therefore has positive \(n\)-dimensional volume.

Recipe 6: if a power is invertible

Given: \(A^k\) is invertible for some positive integer \(k\).
Prove: \(A\) is invertible.

Determinants provide the shortest route:

The left side is nonzero, so \(\det(A)\ne0\). Alternatively, if \(Ax=0\), then \(A^kx=0\); invertibility of \(A^k\) forces \(x=0\), so \(A\) has trivial kernel.

Recipe 7: if a product is invertible

Given: \(A,B\in\mathbb R^{n\times n}\) and \(AB\) is invertible.
Prove: both \(A\) and \(B\) are invertible.

Using determinants,

so neither factor can have zero determinant. Thus both are invertible. The inverse is

This conclusion requires square factors of the same size. Rectangular factorization needs a more careful rank argument.

Recipe 8: parameter values that make a matrix invertible

Determine when

is invertible. Compute

Therefore \(A(k)\) is invertible for every real \(k\) except

At either excluded value, the determinant is zero, rank drops, a nonzero null vector appears, zero becomes an eigenvalue, and all other IMT conditions fail.

Recipe 9: a triangular matrix

For an upper triangular matrix,

the determinant is the product of diagonal entries:

Therefore \(U\) is invertible exactly when every diagonal entry is nonzero. Those diagonal entries are also the eigenvalues, and they occupy the pivot positions. One visual check simultaneously reads the elimination, determinant, and spectrum conditions.

Recipe 10: invertible but ill-conditioned

Let

Then \(\det(C_\varepsilon)=\varepsilon\ne0\), so the matrix is invertible. Its singular values are \(1\) and \(\varepsilon\), giving

As \(\varepsilon\) approaches zero, exact invertibility remains true for every positive \(\varepsilon\), but numerical sensitivity becomes arbitrarily severe. This is the cleanest example of why “invertible” and “well-conditioned” are not synonyms.

Recipe 11: a rectangular Gram matrix

Let \(X\in\mathbb R^{m\times n}\) with \(m\ge n\). Then

The proof uses

This is not the square IMT applied directly to \(X\); it is the IMT applied to the square \(n\times n\) Gram matrix \(X^TX\), together with a kernel identity. This distinction is central in least squares and statistics.

Recipe 12: cancellation

Suppose \(A\) is invertible and \(AB=AC\). Multiplying on the left by \(A^{-1}\) gives

If instead \(BA=CA\), multiply on the right by \(A^{-1}\). Cancellation works because \(A\) is invertible. It is not valid for arbitrary matrices: a singular matrix can erase the nonzero difference \(B-C\).

Self-check problems

Try each problem using only one or two implication bridges before reading the answer.

1. If \(\operatorname{nullity}(A^T)=0\) for a square matrix, what can you conclude about the rows, determinant, and every system \(Ax=b\)?
   Answer: \(A^T\) is injective and invertible; the rows of \(A\) form a basis, \(\det(A)\ne0\), and every \(Ax=b\) has exactly one solution.

2. If the columns of \(A\) span \(\mathbb R^n\), can zero be an eigenvalue?
   Answer: No. Spanning gives full rank and invertibility, so the null space is trivial and zero is absent from the spectrum.

3. If \(\det(A^TA)=0\), what happens to \(\sigma_{\min}(A)\)?
   Answer: Since \(\det(A^TA)=\prod_i\sigma_i^2\), at least one singular value is zero; in particular \(\sigma_{\min}(A)=0\), so \(A\) is singular.

4. If an \(n\times n\) matrix maps one basis to a basis, must it be invertible?
   Answer: Yes. The images of that basis span the codomain and are independent, so the transformation is bijective. Condition 48 uses “every basis,” but even one mapped basis is enough.

5. If \(A\) is invertible, is \(A+A^T\) necessarily invertible?
   Answer: No. Take \(A=\left[\begin{smallmatrix}0&-1\\1&0\end{smallmatrix}\right]\). Then \(A\) is invertible but \(A+A^T=0\). Invertibility is not preserved by matrix addition.