Exact Fits and Zero Residual Variation

When the Fitted Line Has No Residuals

Least squares chooses the line that minimizes the sum of squared residuals. Usually, the fitted values do not equal every observed response, so some residual variation remains. But sometimes all the data points lie exactly on one straight line. Then every residual is zero, and the residual sum of squares is zero as well. The calculation is straightforward; interpreting it requires more care.

In particular, zero residual variation in a dataset does not automatically mean that the underlying process has no random error. It may reflect a deterministic relationship, a small dataset, or values rounded before being recorded. And although the standard residual-based estimate of error variance is then zero, zero is not a usable estimate of a positive noise level for further uncertainty calculations.

The Model and the Exact-Fit Calculation

Write the simple linear regression model as \(Y_i=\beta_0+\beta_1x_i+\varepsilon_i\), where \(x_i\) is the predictor value for observation \(i\), and \(Y_i\) is its response. For the standard variance interpretation, assume the \(x_i\) values are fixed, the errors are independent with mean zero and common variance \(\sigma^2\), and the predictor values are not all equal. Normal errors, \(\varepsilon_i\sim N(0,\sigma^2)\), are an additional assumption when normal-based inference is required; they are not needed to calculate a least-squares line.

The least-squares estimates \(\hat{\beta}_0\) and \(\hat{\beta}_1\) give fitted values \(\hat{Y}_i=\hat{\beta}_0+\hat{\beta}_1x_i\). The residual for observation \(i\) is \(e_i=Y_i-\hat{Y}_i\). The residual sum of squares is \(SSE=\sum_{i=1}^n e_i^2\). With \(n\) observations and an estimated intercept and slope, the residual-based estimate of error variance is \(s^2=SSE/(n-2)\), provided \(n>2\).

The condition \(n>2\) matters. With two observations at distinct predictor values, a line can pass through both points, but there are no residual degrees of freedom left to estimate error variance. A zero residual sum of squares from two points is therefore automatic, not evidence of a noiseless relationship.

Why an Exact Line Gives Zero Residuals

Suppose the observed values satisfy \(Y_i=a+bx_i\) for constants \(a\) and \(b\), with at least two distinct predictor values. The least-squares slope can be written as \(S_{xy}/S_{xx}\), where \(S_{xy}=\sum (x_i-\bar{x})(Y_i-\bar{Y})\) and \(S_{xx}=\sum (x_i-\bar{x})^2\). Substituting the exact-line relationship shows why the calculated slope and intercept recover that line.

Thus \(\hat{Y}_i=a+bx_i=Y_i\) for every observation, so \(e_i=0\) and \(SSE=\sum 0^2=0\). If \(n>2\), dividing by the positive residual degrees of freedom \(n-2\) still gives \(s^2=0\). This establishes the calculation, not the truth of a zero-variance population model.

What Zero Residual Variation Means

The residual variance estimate \(s^2\) summarizes the variation left after fitting the line. When all residuals vanish, the formula has no residual variation to summarize and returns zero. It does not distinguish why the residuals vanished. For example, a process may be deterministic, recorded values may have been rounded, or a small collection of observations may happen to display an exact pattern.

There is also a model-based caution. Under the normal regression model with \(\sigma^2>0\), errors have a continuous distribution. With more than two observations and a full-rank design, the probability that the entire random error vector lands exactly on the fitted line is zero. Thus an exact fit in such a dataset is not an ordinary positive-variance normal-regression outcome. It can indicate that the model or the data-recording description needs scrutiny. It does not justify quietly assuming that \(\sigma^2=0\) while continuing to use formulas that require positive error variance.

If a calculation requires a standard deviation based on \(s^2\), the result here is zero. That is a warning, not evidence that estimated uncertainty has disappeared. In particular, dividing by a standard error calculated as zero is not a valid way to claim overwhelming evidence. The variance estimate has collapsed because the observed residuals contain no variation.

Nearly Exact Fits and Recorded Precision

A dataset can look perfectly linear in a table that displays only a few decimal places, even when the underlying recorded values are not exactly on a line. To decide whether the fit is truly exact, use the values supplied in the problem and compute the residuals. Do not replace small residuals with zero merely because they seem negligible. Conversely, do not claim extra precision beyond the stated data.

For a nonhorizontal exact fit, \(S_{YY}>0\) and \(SSE=0\), so the usual \(R^2\) formula gives 1. For a horizontal exact fit, \(S_{YY}=0\), and the formula is undefined. A second trap is confusing these cases with the two-observation situation: with two distinct predictor values and only two observations, an exact line can always be drawn, but \(n-2=0\), so \(s^2\) cannot be calculated.

Exam Checks and Common Traps

The central distinction is between a property of the observed dataset and a claim about the stochastic process that generated it. An exact fit is a statement about observed residuals. The regression error variance is a model quantity. A zero residual-based estimate signals that this dataset supplies no positive residual variation from which to estimate a positive noise level.