A column whose values are all absent sums to zero, exactly as an all-zero column would — why, and when does it not?

## Skip first, fold second An aggregate that steps over absent values does two things in order. First it discards the holes. Then it folds what remains — adds them up, multiplies them, takes the largest. The order is the whole explanation: if every value in the column was absent, the first step discards everything and the second step is asked to fold **an empty set**. A fold over an empty set still has to return something, and the value chosen is the operation's **identity element**: the value that leaves the fold consistent when the input is split into parts and one part happens to be empty. Adding the total of an empty part to the total of the rest must give the total of the whole, and only zero does that for addition. Only one does it for multiplication. ## What that produces in a report A column where every value is absent sums to **0**, and that 0 is the same 0 a column of genuine zero measurements would produce. This is the one place in the subject where three unrelated zeros become literally the same value on the screen: - a **measured zero** — the reading really was nothing; - a **substituted zero** — somebody wrote it into the holes earlier in the pipeline; - the **zero a sum returns having added nothing at all**. Downstream, nothing distinguishes them. The tile renders 0, the chart draws a bar of height zero, the threshold alert compares 0 against its limit and stays quiet. It is worth saying plainly: the sum is not lying. It is answering the question "what is the total of the values present?" perfectly correctly, and the answer to that question over no values is zero. The mismatch is between that question and the one the reader thinks they asked. ## Not every aggregate does this | Aggregate | Identity over an empty set | What a skipping surface returns when nothing was present | |---|---|---| | sum | 0 | 0 | | product | 1 | 1 | | count of present values | 0 | 0 | | mean | none — it would be a division by zero | absence, or an undefined-arithmetic result | | smallest or largest value | none | absence, an error, or an infinity, depending on the design | The split is not arbitrary. **Sum, product and count have an identity; mean, smallest and largest do not.** That is why the failure is characteristically a total rather than an average: the average over nothing is visibly nothing, while the total over nothing looks like a result. For the smallest and largest value the designs genuinely disagree — some hand back absence, some refuse the call, some return an infinity of the appropriate sign — so do not carry one tool's behaviour to another. The undefined-arithmetic result named above is the pattern the floating-point format itself sets aside for a computation with no answer; whether a tool exposes it, or converts it to its own absence marker, is another point on which designs differ. ## The two settings that change the answer 1. **The default direction of the surface.** Where an aggregate defaults to propagating absence rather than skipping it, a column containing even one hole returns absence, and an all-absent column certainly does. That default is noisier and, for this particular failure, safer. 2. **The minimum-present threshold** — the setting that makes a fold return absence unless at least that many real values took part. Set to one, it converts exactly this case from a confident zero into a visible gap while leaving every other total untouched. Where a surface offers it, it is close to free insurance for any total that is reported rather than consumed internally. Neither setting is the same as the other, and neither is a property of the data. Both are decisions taken at the call. ## Detecting it when you cannot change the call - Report the **count of present values** beside every total. A total of 0 whose present count is 0 is not a total; a total of 0 whose present count is 4,000 is a real result. One extra column in the output separates them for the price of one pass. - Assert a minimum completeness for the columns a report depends on, so a column that becomes entirely absent upstream fails loudly rather than resolving to zero. - Be alert to the case where a **subset of rows** — rows that do exist — happens to have nothing recorded in the column being totalled. The subset is not empty, so a row-count check passes, and the total is still a fold over nothing. - Distrust a zero that appears where a zero was never plausible before. A metric that has never been zero and is suddenly zero is more likely a column that stopped arriving than a business that stopped trading. ## In an interview Say the mechanism — skip first, fold second, and a fold over nothing returns the identity — then say the consequence, that the zero is indistinguishable from a real total, and then name the two things that change it. The extra half-mark is for pointing out that a mean over nothing does *not* come back as zero, because it has no identity to fall back on.