Sometimes the evidence genuinely cannot decide between two explanations — not because it's weak, but because both predict exactly the same observations. Was he quiet from confidence or from fear? Did the policy work, or did the crisis pass? When two theories are observationally identical, no amount of the same data will ever tell them apart. This is underdetermination with teeth.
He sat through the whole meeting and said almost nothing. Afterwards everyone had a reading. He's so sure of himself he didn't need to weigh in, said one. He was terrified — couldn't find a way in, said another. Both watched the same silence; both are certain. And here is the uncomfortable part: watching him sit through a second meeting, and a third, will not settle it. In that room, the behaviour confidence produces and the behaviour fear produces are the same behaviour. The quiet does not carry the label.
We carry a hopeful rule of thumb — more evidence narrows the truth. Usually it does. But some questions have a ceiling built into them. The observations you can gather are consistent with several genuinely different underlying stories, and collecting more of the same observations moves you not one inch, because every one of those stories predicts exactly what you are seeing. The ambiguity is structural, not statistical. Statisticians call such a question unidentifiable, and the mathematics that measures the ceiling is the plain linear algebra of rank and the null space — the difference between we don't know yet and this data can never know.
Behind the silence are three hidden quantities you'd love to know: confidence, fear, and competence. You cannot see them. You can only gather observations — each one some combination of the three. Switch the kinds of evidence on and off and the instrument computes, exactly, how many genuinely independent things your evidence pins down (the rank) against the three you want to know. When rank falls short, drag the slider along the null space and watch the underlying story change wildly while every observation stays frozen.
The kinds of evidence you can gather
The hidden story the evidence is meant to reveal
What the evidence actually sees
Think of a model as a machine that turns hidden causes into visible observations. Each observation you take is a constraint on the unknowns — an equation they must satisfy. Rank is not the number of measurements you pile up; it is the number of those constraints that are genuinely independent — how many distinct questions your evidence actually answers. Take the same measurement twice and you have two readings but one question; the rank stays put. This is why a courtroom can hear a witness say the same thing in ten different ways and learn nothing new: ten sentences, rank one.
When rank is less than the number of unknowns, something exact and unsettling appears: a whole null space of adjustments you can make to the hidden causes that cancel out perfectly in every observation. Add confidence, subtract the same amount of fear, and the demeanor — which sees only their sum — does not flinch. Every point in that space is a different underlying story, and the data ranks them all exactly equally. Unidentifiability is precisely this: a null space that refuses to close. It is not a gap in your evidence that patience will fill. It is a direction the evidence is structurally blind to.
Two questions can both feel undecided and need opposite cures. The diagnosis panel takes a real question and its available evidence, builds the little matrix, and computes which of three worlds you're in. Click through them.
The fork is everything. A weakly identified question has full rank — the evidence can in principle decide it — but only barely, so the answer is buried in noise; here, more of the same data is exactly right, and patience pays. An unidentified question is rank-deficient: no quantity of the same evidence will ever separate the stories, and the only real moves are a different kind of observation or an assumption imposed from outside. Pouring more of the same evidence onto an unidentified question — a hundred more meetings, a thousand more surveys of the treated group — is the most common wasted effort in inquiry, in courtrooms, and in self-knowledge. Diligence aimed down the null space is just motion.
Some questions about people, history, and yourself are unidentifiable with any evidence you'll ever have. Motive against motive; cause against coincidence. Did you leave the job from courage or from cowardice dressed as principle? The behaviour is one; the stories are two; the record you can consult sees only the sum. The honest posture is not to pick the flattering story and call it settled. It is to say plainly: this cannot be known from here — and to notice how much comfort we take in mistaking a null direction for a resolved one.
When you must decide anyway, there are exactly two legitimate moves, and one cheat. You can break the tie by an explicit assumption — an identifying restriction stated out loud (assume the trend would have been flat; assume he'd have spoken if confident) — so that everyone can see the answer rests on it and argue the assumption rather than the data. Or you can hold both stories, carrying the ambiguity honestly. The cheat is to smuggle the assumption in unstated and present a chosen point in the null space as though the evidence had singled it out. This is the same shape as an inverse problem, where many causes fit one effect and you must choose a solution knowing it's a choice; and it is why a Bayesian prior is an assumption made visible — the outside information that picks among observationally equal stories. State it, and it's honest. Hide it, and it's a con.
| Mathematics | Life |
|---|---|
| the unknowns | the competing explanations or hidden causes — confidence, fear, the treatment, the trend |
| the observations | the evidence you can actually gather — the only part of a cause anyone ever sees |
| the rank | how many genuinely independent things the evidence pins down — not how much of it there is |
| the null space | the family of different stories that fit the observations identically |
| unidentified | a question no amount of the same data will ever decide |
| a new kind of observation | the only real fix — or an assumption honestly imposed from outside |
More evidence narrows the truth only along the directions the evidence can see. Down the null space, a library of data and a single glance say exactly the same thing.