Zero Pivots and Elimination Breakdown

What a Pivot Does

In Gaussian elimination, the pivot is the entry used to clear entries below it in the current column. For an \(n\times n\) matrix, the first pivot is the entry in position \((1,1)\). After clearing below it, the next pivot is the entry in position \((2,2)\) of the updated matrix, and the process continues down the diagonal. These are the pivots of the elimination process; after earlier row operations, they need not equal the original diagonal entries.

Suppose the current pivot is \(a_{kk}^{(k)}\), and the entry beneath it in row \(i\) is \(a_{ik}^{(k)}\). When the pivot is nonzero, the elimination multiplier is the quotient \(a_{ik}^{(k)}/a_{kk}^{(k)}\). Subtracting that multiple of row \(k\) from row \(i\) makes the entry in column \(k\) zero. This simple division explains two different concerns: a zero pivot makes that formula unavailable, while a nonzero but very small pivot can make the multiplier large.

These distinctions matter. A zero pivot can occur in a nonsingular matrix, and a small pivot does not, on its own, prove that elimination will be inaccurate. The exact obstruction and the numerical warning have different causes and should not be treated as the same diagnosis.

Why a Zero Pivot Can Block Elimination

Proof. Write \(c\) for the multiple of row \(k\) to subtract from row \(i\). The resulting entry in column \(k\) is \(a_{ik}^{(k)}-c\,a_{kk}^{(k)}\). If \(a_{kk}^{(k)}\ne0\), setting this expression to zero requires \(c=a_{ik}^{(k)}/a_{kk}^{(k)}\), and this value is unique. If \(a_{kk}^{(k)}=0\), the expression is \(a_{ik}^{(k)}\) for every \(c\). When \(a_{ik}^{(k)}\ne0\), it therefore cannot be made zero this way. \(\square\)

The zero-pivot obstruction is specifically about continuing the current column without changing rows. If the pivot is zero but every entry below it in that column is already zero, there is nothing to clear in that column. The zero diagonal entry is still important: it is not a nonzero pivot, and it may prevent a triangular solve from determining a unique solution. Do not label both situations as the same failed division; check whether there is a nonzero entry below the pivot.

There is another zero-pivot situation. If a pivot is zero and all entries below it in its column are zero, no elimination is needed in that column, but the zero diagonal entry can still matter when solving a triangular system. For example, a zero row produced by elimination may signal that the equations are dependent rather than inconsistent; what happens depends on the right-hand side as well as the matrix. The immediate diagnosis should describe the pivot and the entries beneath it, rather than jumping to a conclusion about every associated system.

Why a Small Pivot Raises a Numerical Warning

With a nonzero pivot, elimination is defined in exact arithmetic, however small that pivot is. But the multiplier is a quotient, so its size depends on both the entry being cleared and the pivot. If the pivot is tiny while an entry below it is not, the multiplier can be large. In finite-precision arithmetic, calculations involving large multipliers can magnify the effects of rounding. This is a reason to investigate a small pivot, not a proof that an answer will be inaccurate.

Proof. Since the pivot is nonzero, the multiplier is defined. Taking absolute values gives \(\left|a_{ik}^{(k)}/a_{kk}^{(k)}\right|=|a_{ik}^{(k)}|/|a_{kk}^{(k)}|\). The numerator is at least \(c\) and the positive denominator is at most \(\varepsilon\), so the quotient is at least \(c/\varepsilon\). \(\square\)

Small Relative to What?

A pivot's absolute size is not enough to assess a warning. The entries being processed also have a scale. For instance, \(\tfrac{1}{10000}\) is small compared with \(1\), as in \(\mathbf{C}\), where it leads to a multiplier of \(10000\). But a small pivot can occur without producing a large multiplier if the entries below it are also small or zero.

Scaling also changes the numerical meaning of an absolute threshold. Multiplying a row and its corresponding right-hand-side entry by a nonzero scalar leaves the set of solutions unchanged, but changes the magnitudes appearing in that row. Therefore, a rule such as “treat every pivot below a fixed number as bad” can give different warnings for differently scaled versions of the same equations. In practice, a small-pivot warning is assessed relative to the entries under consideration, with attention to how the system is scaled and to the precision of the computation.

Common Mistake: Treating Both Cases as the Same

Wrong calculation: In the matrix \(\mathbf{A}=\begin{bmatrix}0&1\\1&0\end{bmatrix}\), one might write the multiplier as \(1/0\) and continue as though this were an ordinary number, or conclude from the failed division that \(\mathbf{A}\) has no inverse. Neither step is valid. Division by zero is undefined, and the matrix is invertible, as the check \(\mathbf{A}^2=\mathbf{I}_2\) shows.

Correction: State what has actually failed: no-pivot elimination cannot clear the first column using the first row. Separately check invertibility or the behavior of the system; do not infer singularity merely from that blocked step. For a nonzero small pivot, use the actual multiplier formula and compare the resulting multiplier and pivot with the scale of the active entries. A numerical warning calls for care, not an automatic declaration of failure.

These checks keep three questions separate: Is the current pivot exactly zero? If so, is there a nonzero entry below it that the current row must clear? If the pivot is nonzero but small, does its size relative to the entries being eliminated create a large multiplier? Answering them in that order distinguishes an exact obstacle from a numerical concern without overclaiming what either one implies.