belief & evidence · metaphor 82 of 100

Every inference has a
condition number.

Two people revise their conclusions when the facts wobble. One shifts a little; the other swings from devotion to despair on the same small news. The difference is not temperament but geometry: every inference has a condition number — how violently its output responds to error in its inputs — and you can compute yours.

A ship takes bearings on two lighthouses. If the lighthouses stand at right angles from the ship, a small error in either bearing barely moves the fix. If they stand nearly in line, the same small error throws the fix miles up the coast. Nothing about the instruments changed — only the angle between the constraints. The compass is as good as it ever was; the geometry of the evidence has quietly turned it into a rumor mill.

Plans, diagnoses, self-assessments: all are fixes computed from bearings. You believe you are good at your work because two colleagues said so; you believe the company is healthy because two indicators agree. Every such conclusion is an intersection of sight-lines, and most people never ask at what angle their evidence crosses. Drag the lighthouses below and watch a confident conclusion smear into a long red maybe.

Relabel the bearings
lighthouse A lighthouse B true position 400 resampled fixes + 2σ ellipse

drag a lighthouse · or focus the chart and use keys 1 / 2 / 3 + arrows · a 10 km stretch of coast

crossing angle A∠B
condition number κ
1° of bearing error →
cloud axis ratio ≈ κ
Measurement error · ε · the honest wobble
±1.00°
↤ steady hand · ±0.25° shaky world · ±5° ↦
Watch what happens
Every red dot is an honest fix: both bearings re-measured with Gaussian error ε, lines intersected, dot plotted. The cloud is your uncertainty.
κ = σmax σmin The ratio of the largest to smallest singular value of the bearing system. κ multiplies your error: worst-case output error ≈ κ × best-case output error, per unit of input wobble. For two bearings at equal range crossing at angle φ, κ = cot(φ/2) — exactly 1 at 90°, about 7.6 at 15°, unbounded as the lines close.

nothing is faked: κ from the singular values of the live bearing matrix (each constraint weighted by its range, so κ equals the noise ellipse's expected axis ratio); each dot a real weighted-least-squares fix under resampled Gaussian noise; the ellipse from the sample covariance of those dots.

The angle of your evidence

Independence is a geometry, not a headcount.

Each lighthouse contributes one constraint: the truth lies somewhere along this line. A conclusion needs the lines to cross, and everything depends on how they cross. At 90° the constraints are perpendicular — each one forbids exactly what the other permits — and error in one barely leaks into the answer. As the angle closes, the two lines say more and more nearly the same thing, and the intersection becomes an argument between near-duplicates about a place neither can see. Two correlated sources are one bearing wearing two coats.

This is the geometry of the echo chamber, and the instrument shows its signature: the cloud of possible conclusions stretches into a long thin ellipse, pointed at nothing. Along the short axis — the direction both sources genuinely constrain — the echo chamber is exquisitely precise. Along the long axis, it knows almost nothing, while feeling no less certain. The ratio of those two axes is κ itself: agreement between your sources measures their overlap, not the world.

What to try

Sixty seconds at the chart table.

Drag the lighthouses toward each other until the crossing angle reads about 15°, and watch the tight dot of fixes smear into a streak — the same ±1° of bearing error, amplified roughly eightfold along one direction. Then press the third lighthouse: one genuinely independent bearing, and κ collapses; the least-squares fix snaps back to a dot even though the first two sources are as redundant as ever.

Now raise ε — make the whole world shakier — and notice that the well-conditioned fix barely cares. At 90°, quintupling the measurement error merely fattens a small dot; at 10°, even a steady hand cannot save you. Robustness is bought at the crossing angle, not at the instrument. Better lighthouses are worth far less than better-placed ones. Finally, use the relabel buttons: the geometry is indifferent to whether the bearings are lights on a coast, friends' opinions, business dashboards, or a symptom and a test.

Brittle plans, brittle selves

What small error would swing your conclusion completely?

Consider a plan whose success requires two nearly-redundant things to be exactly right — the launch works if the market grows and the closely-tied partner deal lands. Those are two bearings taken from almost the same spot; the plan's fix has an enormous κ, and a sliver of bad news anywhere sends the projected future sliding miles along the ellipse. Or consider a self-esteem triangulated from two correlated mirrors — two friends from the same circle, two metrics from the same boss. Same pathology: the person who swings from devotion to despair on one stray remark is not oversensitive; they are ill-conditioned, running their identity through nearly-parallel constraints.

This yields a diagnostic question you can ask of any conclusion you hold: what small error in my inputs would swing this completely? Not "am I sure of my sources" — the ship's officers were sure of both lighthouses — but "if each source is off by its honest ±ε, how far does my answer travel?" A belief that survives a wobble in every input is well-conditioned. A belief that dies of a rounding error was never a fix; it was a coincidence of sight-lines.

Conditioning your life

Diversify the angles, not the number.

A third nearly-parallel lighthouse buys you almost nothing — ten pundits from one school are still one bearing — while a single line from a genuinely different direction transforms the problem. So diversify the angles of your evidence, not its quantity: one outsider, one base rate, one person who owes you nothing. Prefer conclusions that survive ±ε on every input, and treat "all my sources agree" as a question — agree from where?

And when κ is irreducibly huge — some questions are like that; reality does not owe you perpendicular evidence — the discipline is not to pretend otherwise. Hold the fix loosely and say so: report the ellipse, not the dot. "It's probably here, but a hair of error moves it far, and in this particular direction" is not weakness. It is the only honest sentence a navigator on that coast can say.

The mapping

Mathematics ↔ life.

MathematicsLife
the bearingsYour sources of evidence — each one a line the truth must lie on, never a point.
angle between themTheir real independence: how differently they see, not how many they are.
the fixThe conclusion, the plan, the self-image — always an intersection, never an observation.
error εThe honest wobble in every input; no bearing was ever exact.
condition number κHow violently the conclusion swings per unit of wobble — the brittleness of the inference itself.
the third lighthouseOne genuinely different kind of evidence, worth more than ten echoes.

Where the metaphor tears

Three honest failures.

κ measures sensitivity, not truth.

A well-conditioned inference can be confidently, stably wrong: if both lighthouses are mischarted the same way, the fix is a tight little dot in the wrong sea. Conditioning is about noise; calibration is about bias. Perpendicular evidence protects you from wobble, not from a shared delusion — you need both disciplines, and this instrument shows only one.

Your inputs are not instruments.

A bearing is a separable measurement with a knowable ±ε. A friend's opinion of you is not: it entangles with your other sources, shifts when observed, and carries no error bar. The formalism idealizes — treating testimony, symptoms, and dashboards as clean lines with Gaussian wobble. Use κ as a question to ask, not a number to compute about people.

Some brittleness is chosen rightly.

High-κ inferences from nearly-parallel evidence are sometimes all reality offers — the rare disease with one ambiguous test, the crime with a single witness. The counsel is not "never be there"; sometimes there is where the truth lives. The counsel is to know you're there: to see the long thin ellipse for what it is and price your confidence accordingly.