Scalars, Units, and Coordinate Systems

A stored number is the smallest part of a feature. Around it sit the unit it is expressed in, the origin its differences are measured from, the process that produced it, and the range of decisions it is accurate enough to support. Strip any one of those layers away and the number survives intact. Its meaning does not.

Pounds mistaken for kilograms is a documented category of medication error, and the Pennsylvania Patient Safety Authority counted it. It analysed 479 medication-error reports involving patient weights, submitted between June 2004 and the end of November 2008. 129 of them — 26.9% — were classified as "Confusion between pounds versus kilograms". 322 of the 479 events, 67.2%, reached the patient, and 43.4% were wrong dose or overdosage. The Authority stated the pattern in its own words: “In fact, more than 25% of the 479 reports mention breakdowns that occurred when the patient’s weight, measured in pounds or kilograms, was erroneously documented as the patient’s weight in kilograms or pounds, respectively.”

The whole mechanism fits inside one stored value. NCC MERP's 2018 weighing recommendation gives it as “a weight of 220 lb. documented as 220 kg instead of 100 kg”. Nothing about the number 220 is wrong. It is a correct reading in one unit and a greater-than-two-fold dosing error in the other, and the column that holds it cannot tell you which. So the remedy is not a better model of the data. It is a change to the measurement process itself: metric-only documentation, and scales locked out of pounds.

A scalar is a single numerical value, and its meaning depends on how it was measured. The value 20 could be degrees Celsius, kilograms, milliseconds, or a count of previous purchases. The value 220, as the 479 Pennsylvania reports show, can mean a 100 kg patient or a dose calculated for someone twice that mass. Machine-learning pipelines often strip that context away. A warehouse column may hold clean numbers while mixing currencies, time zones, sensor calibrations, and definitions from different teams. Mathematics begins before modeling, at the moment we decide which quantities are comparable. Units, origins, and scales determine whether addition, subtraction, ratios, and distances make sense at all.

These incidents are mathematical before they are algorithmic, and the fullest documented example is an aircraft.

On 23 July 1983 Air Canada Flight 143 ran out of fuel over Ontario and landed without engines at Gimli, Manitoba. The aircraft was a Boeing 767, registered C-GAUN. The board of inquiry under Mr Justice George H. Lockwood reported in April 1985 and reconstructed the arithmetic step by step.

Drip-stick readings of 62 and 64 cm gave 3,758 + 3,924 = 7,682 litres already on board. The crew multiplied 7,682 by the fuelling conversion factor 1.77 and obtained 13,597, which they treated as kilograms. The flight plan required 22,300 kg, so 22,300 − 13,597 = 8,703, and 8,703 / 1.77 = 4,916 litres to be added, rounded at the fueller's suggestion to the 5,000 litres actually boarded in Montreal. The commissioner's finding on that step is one sentence: “Multiplying by 1.77 gave pounds, but everyone involved thought they were kilograms.”

The correct calculation divides by 2.2 as well. 7,682 x 1.77 / 2.2 = 6,180 kg on board. 22,300 − 6,180 = 16,120 kg still required, which called for 20,036 litres rather than 5,000. The aircraft was cruising at 41,000 feet when the low-fuel-pressure warnings began, and both engines failed as it descended through 35,000 feet. Every multiplication was performed correctly. The factor 1.77 simply carried a unit that nobody read.

A point on a map can be described by latitude and longitude, by local east–north coordinates, or by a three-dimensional vector from Earth’s center. The location stays the same while its coordinates change.

Australia performed the demonstration at continental scale. Geoscience Australia launched GDA2020 on 15 December 2017 and announced it plainly: “Australia's coordinates have officially moved 1.8 metres northeast, with Minister for Resources and Northern Australia Matt Canavan today launching Australia's new coordinate reference frame, the Geocentric Datum of Australia 2020 (GDA2020).” Nothing on the ground moved relative to anything else. The datum was re-referenced because the Australian plate travels roughly seven centimetres a year, which had left the 1994 datum 1.6 m out of sync with satellite positioning. GDA2020 is realised through the coordinates of the Australian Fiducial Network, and UN-GGIM records that those coordinates are given “in the International Terrestrial Reference Frame 2014 (ITRF2014) at the Reference Epoch”, the reference epoch being 1 January 2020. A fence that had one pair of numbers in December 2017 has another pair now. It is the same fence.

The same principle appears throughout ML. A document can be represented by word counts, latent topics, or an embedding. A rotation of an embedding space changes individual coordinates but may preserve angles and distances. Coordinate-aware reasoning prevents false conclusions from raw parameter values. A large coefficient can become small after rescaling while the represented prediction function stays exactly as it was. That is the 1.8 metre version of the same fact.

All three alter numbers, and they preserve different relationships. Unit conversion multiplies by a known constant, so ratios and ordering survive, provided the label travels with the value. Centering subtracts a reference such as a mean or a baseline: it moves the origin, preserves every pairwise difference, and gives zero a new meaning. Standardization centers and then divides by a scale estimate, producing dimensionless coordinates. Unlike the other two, it changes the relative distances between features. That is exactly why it changes what a distance-based model sees.

That last claim has been measured. Wongoutong ran k-means on ten real datasets and published the results in PLOS ONE in 2024. On the Penguins data — 4 features in different units, 342 objects, 3 classes — accuracy was 0.5614 on the raw data and 0.9620 after Z-score standardization. Precision moved from 0.4949 to 0.9578. A chi-square homogeneity test against the true grouping was rejected at p<0.001 for the raw data, and not rejected, p=0.925, after standardization. On the datasets whose features already shared a single unit, scaling changed almost nothing. The abstract states the finding: “The results of an experimental study show that, for features with different units, scaling them before k-means clustering provided better accuracy, precision, recall, and F-score values than when using the raw data.”

The effect is not confined to clustering. In Applied Soft Computing in 2023, de Amorim and colleagues compared 5 scaling techniques across 20 classification algorithms and 82 public datasets, and found the gap between the best and the worst scaling technique to be relevant and statistically significant in most cases. Scaling is not a formatting step applied after the modeling decisions are made. On the same 342 penguins it is worth more than a third of the accuracy.

Coefficients, distances, and feature importance are worth auditing before they are interpreted. Name the quantity, and the entity and time grain it is measured at. Record the unit, including currencies, time zones, denominators, and implicit scaling factors. Identify the origin, and decide whether zero means absence, a baseline, or an arbitrary reference. Trace every transformation: centering, standardization, clipping, logarithms, and the statistics they were fitted on. Then test invariance, and ask which conclusions ought to survive an equivalent change of unit or coordinate.

Step three is the one that gets skipped, and Hubble is the definitive example of what skipping it costs. The 2.4 m primary mirror was polished with great precision to the wrong shape, leaving about 256 nm of spherical aberration. NASA's optical systems failure board, chaired by Lew Allen, reported in November 1990 that the aberration came from a measurement reference nobody re-checked: “It has been established that the field lens was approximately 1.3 mm too far from the lower mirror of the RNC, which was used to figure the primary mirror.” The displacement was traced to an interferometer that focused on the end cap of a metering rod whose anti-reflective coating had flaked away. The board found that the measured 1.3 mm error “accounts in detail for the amount and character of the observed image blurring”, and that “No verification of the reflective null corrector's dimensions was carried out by Perkin-Elmer after the original assembly”. Lallo recorded the same cause two decades later: “a 1.3 mm spacing error in the reflective null corrector device used to optically test the mirror while being figured”.

Precision measured against an unverified origin is precision in the wrong direction. It is also the failure mode hardest to detect from the data alone, because every reading is beautifully consistent with every other.