Do Non Square Matrices Have Determinants

You stare at a 3×5 matrix. You need the determinant. Your fingers hover over the keyboard, ready to type det(A) into MATLAB, NumPy, or your calculator of choice.

It errors out. Or worse — it gives you a number that means absolutely nothing.

Here’s the short version: non-square matrices do not have determinants. Full stop. The determinant is simply not defined for rectangular arrays.

But that answer feels unsatisfying, doesn’t it? Even so, because the reason* you wanted a determinant — checking invertibility, measuring volume scaling, finding eigenvalues — those problems don’t vanish just because your matrix isn't square. You still need to solve them.

So let’s talk about what the determinant actually is, why the definition breaks for rectangles, and — crucially — what you reach for instead.

What Is a Determinant (and Why It Demands a Square)

Most introductions define the determinant as a special number calculated from a square matrix. That’s true, but it’s a what*, not a why.

Geometrically, the determinant of an n×n matrix is the signed volume of the n-dimensional parallelepiped spanned by its column vectors. On the flip side, in 3D, the volume of a skewed box. In 2D, that’s the area of a parallelogram. In n dimensions, the hypervolume.

That geometric picture requires* the same number of vectors as dimensions. Input dimension equals output dimension. Worth adding: you can’t define a 2D area with three vectors living in 3D space — not without projecting first. You can’t span a 3D volume with only two vectors. The mapping goes from ℝⁿ to ℝⁿ. That’s why the matrix must be square.

Here's a detail that's worth remembering.

Algebraically, the determinant is the unique (up to scale) multilinear, alternating map from n vectors in ℝⁿ to a scalar. The alternating property — swapping two columns flips the sign — only makes sense when you have exactly n columns in an n-dimensional space.

The Leibniz Formula Perspective

If you’ve seen the Leibniz formula:

$\det(A) = \sum_{\sigma \in S_n} \text{sgn}(\sigma) \prod_{i=1}^n a_{i,\sigma(i)}$

The sum runs over permutations of {1, ...Because of that, , n}. terms. Plus, a 3×5 matrix has 3 rows and 5 columns. There are n! On top of that, each term picks exactly one entry from each row and each column. On the flip side, that “one per row, one per column” constraint is baked into the definition. You can’t pick one entry from each of the 5 columns while only using 3 rows without repeating rows — which the permutation structure forbids.

The Laplace Expansion Perspective

Cofactor expansion along a row or column? Same issue. Plus, you reduce an n×n determinant to a sum of (n-1)×(n-1) determinants. Which means the recursion bottoms out at 1×1. A rectangular matrix never hits that base case cleanly.

Why It Matters: What the Determinant Does*

People don’t usually want the determinant for its own sake. They want what it tells* them. For square matrices, that’s a lot:

• Invertibility: det(A) ≠ 0 ⇔ A is invertible.
• Volume scaling: |det(A)| is the factor by which A stretches volumes.
• Orientation: sign of det(A) tells you if A flips space.
• Eigenvalues: det(A - λI) = 0 is the characteristic polynomial.
• Linear independence: columns are independent ⇔ det ≠ 0.
• Change of variables: the Jacobian determinant in integration.

When your matrix isn’t square, every single one of these questions still exists*. You just can’t answer them with a single scalar called “the determinant.”

How It Works: The Tools That Replace the Determinant

This is where the real work happens. Depending on which property you cared about, you reach for a different tool.

Rank and Singular Values: The Invertibility/Independence Replacement

For a square matrix, det ≠ 0 means full rank. For an m×n matrix A, rank is the direct generalization. It tells you the dimension of the column space (and row space — they’re equal).

But rank is an integer. In real terms, it’s coarse. It doesn’t tell you how close* you are to losing rank.

That’s where Singular Value Decomposition (SVD) comes in. Any real m×n matrix factors as:

$A = U \Sigma V^T$

where U is m×m orthogonal, V is n×n orthogonal, and Σ is m×n diagonal with non-negative entries σ₁ ≥ σ₂ ≥ ... ≥ σᵣ > 0 (r = rank).

The singular values σᵢ are the square roots of the eigenvalues of AᵀA (or AAᵀ). The product of the non-zero singular values? They generalize the magnitude of eigenvalues. That’s the pseudo-determinant — the volume of the parallelepiped spanned by the columns in their own subspace.

If you need a condition number, you use σ₁/σᵣ. If you need to know “is this matrix numerically singular?”, you check if σᵣ is tiny relative to σ₁. This is the workhorse replacement for “check the determinant.

The Gram Determinant: Volume in the Subspace

Say A is m×n with m > n (tall matrix, more rows than columns). Its columns are n vectors in ℝᵐ. That's why they span an n-dimensional subspace (assuming full column rank). What’s the n-dimensional volume of the parallelepiped they form?

You can’t use det(A) — A isn’t square. But you can use the Gram determinant:

$\text{Vol}_n = \sqrt{\det(A^T A)}$

AᵀA is n×n, square, symmetric, positive (semi)definite. Its determinant is the square of the n-volume. This works for any m×n matrix with m ≥ n.

If m < n (wide matrix), use √det(AAᵀ) instead — that gives the m-volume spanned by the rows.

This shows up constantly in:

• Least squares (the normal equations involve A

The Gram Determinant in Practice

When the columns of A (an m × n matrix with m ≥ n) are linearly independent, the n‑dimensional volume of the parallelepiped they generate is given by

[ \operatorname{Vol}_n = \sqrt{\det(A^{!T}A)} . ]

Because AᵀA is always square and positive‑definite under full‑column‑rank conditions, its determinant is well defined and non‑zero. This quantity appears in a host of places where “volume” matters even though the original transformation lives in a higher‑dimensional space:

• Least‑squares regression.
  The normal‑equation matrix ( (X^{!T}X) ) is precisely the Gram matrix of the design matrix (X). Its determinant governs the precision of the estimated coefficients: a small value signals that the fitted hyper‑plane is “flat” in some direction, meaning the data are nearly collinear and the estimates become unstable.

• Multivariate normal densities.
  The normalising constant of a k‑dimensional Gaussian with covariance matrix ( \Sigma ) involves (|\Sigma|^{-1/2}). When (\Sigma) is estimated from data as ( \hat\Sigma = \frac{1}{N}X^{!T}X ), the same Gram determinant shows up, linking the likelihood to the volume of the ellipsoid defined by the data cloud.

• Geometric statistics.
  In topics such as shape analysis or principal component analysis, the Gram determinant quantifies how much “spread” the data occupy in their intrinsic subspace, independent of any embedding.

Because the determinant of a Gram matrix is itself a determinant, it inherits all the algebraic niceties (multiplicativity, invariance under orthogonal changes of basis) while being applicable to rectangular matrices.

Pseudo‑determinant: The Product of Non‑Zero Singular Values

If A is rank‑deficient, the ordinary determinant of (A^{!T}A) collapses to zero, losing any information about the subspace that is occupied. The pseudo‑determinant restores a meaningful scalar:

[ \operatorname{pdet}(A) ;=; \prod_{i=1}^{r} \sigma_i, ]

where ( \sigma_1,\dots,\sigma_r ) are the non‑zero singular values of A and ( r = \operatorname{rank}(A) ). Plus, in other words, it is the product of all singular values except the zeros. This scalar plays the same role as the ordinary determinant for volume when the transformation squashes the space onto a lower‑dimensional subspace: it measures the effective* r-dimensional volume of the image of A.

The pseudo‑determinant appears in:

• Determinant‑based regularisation for covariance estimation when the sample covariance is singular.
• Volume‑preserving transformations in graphics and robotics, where one wants to know how much a mapping expands or contracts an r-dimensional object even if it lives in a higher‑dimensional ambient space.
• Information criteria (e.g., Akaike’s Bayesian Information Criterion) that involve log‑determinants of covariance matrices; the pseudo‑determinant provides a well‑defined substitute when the matrix is not full rank.

Conditioning and Stability: The Ratio of Singular Values

A more refined notion of “how close to singular” than a simple rank test comes from the condition number built from singular values:

For more on this topic, read our article on the coldest layer of the atmosphere or check out ethanol fermentation and lactic acid fermentation.

[ \kappa(A) ;=; \frac{\sigma_{\max}}{\sigma_{\min}} \quad (\text{using only the non‑zero }\sigma_i). ]

When A is square and invertible, this ratio coincides with the usual condition number derived from eigenvalues. For rectangular matrices it tells you how sensitive solutions of linear systems (or least‑squares problems) are to perturbations in the data. A large (\kappa) signals an ill‑conditioned problem, regardless of whether the matrix is square.

Because singular values are invariant under orthogonal transformations, they provide a coordinate‑free measure of anisotropy. This invariance makes them ideal for:

• Pre‑conditioning strategies in iterative solvers.
• Feature scaling in machine learning, where one often rescales columns of a data matrix to equalize their singular values before applying algorithms that assume isotropic input.

Putting It All Together

The determinant, as a single scalar, is a convenient shorthand that works only when the linear map is a bijection on a finite‑dimensional vector space. When the map lives in a higher‑dimensional domain or collapses dimensions, we replace that shorthand with a toolbox:

The table captures the logical progression from a single scalar that works only for bijective maps to a suite of tools that retain the same geometric intuition for arbitrary linear operators. Small thing, real impact.

From Singular Values to Log‑Pseudo‑Determinant

In many statistical and machine‑learning contexts we need a smooth* scalar that behaves like a log‑determinant even when the matrix is rank‑deficient. The pseudo‑determinant provides the raw product of the non‑zero singular values, but taking its logarithm yields a quantity that is both additive and numerically stable:

[ \operatorname{ldp}(A) ;:=; \log\operatorname{pdet}(A)=\sum_{i=1}^{r}\log\sigma_i . ]

Because each (\sigma_i) is strictly positive, (\operatorname{ldp}(A)) is well‑defined for any matrix of full column rank, and it reduces to the familiar (\log|\det(A)|) when (A) is square and invertible. This identity is the cornerstone of algorithms that compute Gaussian‑likelihoods for singular covariance matrices, regularized Gaussian processes, and variational inference schemes that assume a low‑rank latent structure.

A practical advantage of (\operatorname{ldp}(A)) is that it can be estimated without explicitly forming the singular values. Here's a good example: the Hutchinson trace estimator combined with stochastic trace‑squared evaluations yields an unbiased approximation of (\sum_{i=1}^{r}\log\sigma_i) with only matrix‑vector products, a boon when (A) is huge or when forming the full SVD would be prohibitive.

Conditioning in Practice: Scaling and Pre‑conditioning

The ratio (\kappa(A)=\sigma_{\max}/\sigma_{\min}) is invariant under orthogonal transformations, which means we can always rotate the coordinate system without altering its value. This invariance is exploited in two complementary ways:

1. Column‑wise scaling – Multiplying each column of (A) by a positive factor (c_j) scales the corresponding singular values by the same factor. By choosing (c_j) to equalize the magnitudes of the resulting singular values, we can dramatically reduce (\kappa(A)). In practice, this is achieved by a simple diagonal pre‑conditioning matrix (D) such that ( \tilde A = A D) has a more favorable singular‑value spectrum. This technique is ubiquitous in least‑squares solvers, where a diagonal scaling step precedes the application of conjugate‑gradient or LSQR.

2. Row‑wise orthogonalization – QR factorizations with column pivoting automatically select a permutation of the columns that maximizes the magnitude of the leading diagonal entries of (R). Since the diagonal of (R) consists of the first (r) singular values of (A) (up to a constant factor), the pivoting strategy yields a matrix whose condition number is bounded by a modest constant times the original (\kappa(A)). This is why pivoted QR is often preferred to plain QR when the matrix is ill‑conditioned but still full rank.

These conditioning tricks are not merely theoretical; they translate directly into faster convergence of iterative solvers, more accurate gradient estimates in optimization, and reduced numerical drift in scientific simulations.

Beyond Linear Algebra: Information Geometry and Volume Preservation

When a linear map is used to define a change of variables in a probability density, the Jacobian determinant appears in the change‑of‑variables formula. If the map is not invertible, the density is defined only on a lower‑dimensional manifold, and the appropriate Jacobian is precisely the pseudo‑determinant. This observation underlies information‑geometric* notions of volume for statistical manifolds that are intrinsically low‑dimensional, such as exponential families with redundant sufficient statistics.

This is the kind of thing that separates good results from great ones.

In robotics and computer graphics, a transformation matrix that projects a three‑dimensional object onto a plane is used to simulate shadows, depth‑of‑field effects, or to perform view‑dependent rendering. Even though the transformation is singular, the pseudo‑determinant tells us how much area* (or volume* in higher dimensions) of an object is preserved under the projection. Artists and engineers can therefore quantify “how much distortion” a given projection

This insight proves invaluable in real-time rendering pipelines, where computational efficiency is key. By precomputing the MPD of projection matrices, developers can dynamically adjust anti-aliasing techniques or texture resolution to compensate for geometric compression, ensuring that distant objects retain perceptual clarity without overtaxing GPU resources.

Machine Learning and Dimensionality Reduction

In machine learning, the MPD emerges as a linchpin in dimensionality reduction techniques like PCA and kernel PCA. And when data lies on a low-dimensional manifold within a high-dimensional space, the covariance matrix often becomes rank-deficient. Now, traditional determinants collapse to zero, obscuring the true geometric structure. The MPD, however, captures the intrinsic variance by multiplying only the non-zero eigenvalues.

and regression, as well as in model selection criteria that must account for the effective rank of the covariance rather than the ambient dimensionality.

In Gaussian process regression, for instance, the log‑marginal likelihood involves the determinant of the kernel matrix. In practice, when the kernel induces a degenerate covariance structure—as frequently happens with stationary kernels evaluated at closely spaced inputs—the standard determinant vanishes, rendering the likelihood undefined. Replacing it with the pseudo‑determinant restores a meaningful measure of model evidence, allowing practitioners to compare covariance functions and tune hyperparameters on datasets where the feature dimension exceeds the effective sample size.

Similarly, in Gaussian mixture models with singular component covariances, the MPD provides a principled way to evaluate the contribution of each cluster to the overall likelihood, even when some components collapse onto lower‑dimensional subspaces. This capability is particularly relevant in anomaly detection, where normal‑behavior clusters may occupy thin manifolds while outliers populate the surrounding high‑dimensional space.

Beyond these domains, the pseudo‑determinant plays a role in control theory and signal processing. When analyzing the observability or controllability of a linear dynamical system whose state matrix has a non‑trivial null space, the MPD of the associated Gramian matrices quantifies the effective* energy that can be injected into or extracted from the system. Engineers designing reliable controllers for underactuated robots or vibration‑suppressed structures rely on such quantities to assess stability margins without being misled by the artificial singularity of the full‑dimensional determinant.

In information theory, the differential entropy of a singular Gaussian distribution—formalized as a degenerate multivariate normal supported on an affine subspace—is given by the logarithm of the pseudo‑determinant of its covariance matrix, plus a term accounting for the dimension of the support. This generalization ensures that entropy‑based measures such as mutual information and KL divergence remain well defined even when the underlying distributions live on lower‑dimensional manifolds, a scenario that arises routinely in compressed sensing, wireless communications with rank‑deficient channel matrices, and diffusion‑based generative models.

Conclusion

The pseudo‑determinant, though a simple algebraic refinement of the classical determinant, unlocks a remarkably wide range of applications. From rescuing the condition number analysis of rank‑deficient matrices in numerical linear algebra, to quantifying geometric distortion in computer graphics, to enabling principled probabilistic inference in machine learning and signal processing, it consistently reveals the intrinsic structure that the ordinary determinant conceals. Still, as datasets grow ever larger and models grow ever more expressive, the prevalence of rank‑deficient and near‑singular matrices will only increase. The pseudo‑determinant offers a unified, numerically stable lens through which to view the geometry of such matrices, making it an indispensable tool in the modern computational scientist's arsenal.