State the norm axioms
A norm assigns a nonnegative magnitude to a vector, vanishes only at zero, scales by the absolute value of a scalar multiplier, and satisfies the triangle inequality. These requirements constrain what can be called a norm. The sum of squared coordinates is useful but is the square of the Euclidean norm, not itself a norm. Doubling a vector multiplies that squared sum by four rather than two. When reading a function name or loss definition, check whether the returned quantity is a length or a squared length. Similar terminology does not make the underlying mathematical objects interchangeable.
Calculate three common norms
For x=(3,−4), the one-norm is seven, the two-norm is five, and the infinity norm is four. They measure the same vector while emphasizing different features: total absolute component magnitude, Euclidean length, or the largest absolute coordinate. None of these numbers alone identifies the universally best interpretation of an error. Comparing a one-norm from one implementation with an infinity norm from another does not establish an improvement. Use a common norm, compatible units, and the same dimension when comparing results. The norm is part of the measurement definition, not a harmless label added after a calculation.
Measure distance through differences
A norm defines a distance between vectors by measuring their difference. The Euclidean distance from (3,−4) to the origin is five, but equal magnitudes do not imply nearby points. The vectors (3,4) and (−3,−4) both have norm five and lie ten units apart. Distinguish measuring from the origin from comparing two observations. Feature units matter too: converting a height coordinate from meters to centimeters changes its contribution to an unadjusted distance. Scaling choices therefore encode modeling judgments. A distance-based method does not become unit-independent simply because its formula satisfies the mathematical axioms of a norm.
Use an explicit error criterion
Numerical error can be measured by taking a norm of the error vector. Absolute and relative errors answer different questions, and a small absolute error can be large relative to a tiny true value. Define how zero denominators are handled. The intended guarantee also matters: bounding every coordinate suggests a maximum-coordinate measure, while an aggregate energy-like quantity suggests a two-norm. A stopping threshold without its scale, units, and expression is difficult to interpret. Report the criterion as part of the algorithm’s contract so that results from different data sets are not compared under silently different definitions of acceptable error.
Matrix norms need their own definition
A matrix can be measured using an induced operator norm or a different construction such as the Frobenius norm. For diag(3,4), the operator norm induced by Euclidean vector length is four, whereas the Frobenius norm is five. Flattening all entries into one vector and measuring them does not generally measure the matrix’s maximum amplification of an input vector. When a condition number or error bound contains a matrix norm, identify which norm is intended. Different values may reflect different questions rather than an arithmetic mistake. Keep vector magnitude, matrix entry magnitude, and transformation amplification distinct throughout the argument.
Compare maximum and accumulated error
For error (0.1,0.1,0.1,0.1), the infinity norm is 0.1, the one-norm is 0.4, and the two-norm is 0.2. Increasing the number of equally sized error components increases aggregate measures without increasing the largest component. When comparing data of different dimensions, distinguish a change in individual accuracy from a change caused by the aggregation definition. Report any normalization by dimension explicitly instead of silently changing the criterion.