maps & meaning · metaphor 80 of 100

Most coefficients are zero

Of the thousand things that could matter, almost all have coefficient zero. Essentialism gets a bad name, but the empirical fact underneath it is real: most causes contribute nothing, and understanding is largely the search for the few that don't.

A doctor faces a patient with a hundred symptoms. An analyst opens a spreadsheet a thousand columns wide. A person lies awake asking why their life feels wrong, and can name forty plausible reasons. All three are drowning in candidate causes — and all three are saved, when they are saved, by the same quiet empirical regularity: most of the coefficients are zero. The world is mostly sparse. A handful of factors carry a phenomenon and the rest are rounding error.

The art of insight is finding the few that carry the signal — and, harder, resisting the mind's hunger to keep every plausible cause alive, each defended by an anxious "but it might matter." The instrument below lets you build a phenomenon out of many candidate features, decide how many of them really matter, and then watch an honest method try to recover the few and zero the rest. It works beautifully — right up until the world stops being sparse, at which point the same method starts telling comforting lies. Both halves are shown, not hidden.

01 · the instrument

The spectrum of sparsity

Fifty candidate features could explain the outcome. Set how many actually do — from 2 in 50 to 40 in 50 — and watch a real lasso try to recover them. Under true sparsity it catches the few and zeros the rest cleanly. As the truth turns dense, recovery frays and no small story suffices.

weight of each candidate — truth vs. what lasso recovered sparse
true weight (the real cause)
recovered — a real cause, caught
false positive / real cause dropped
3 of 50
sparse · a few real causesdense · many
0.10
keep everythingdemand a short story
0.18
cleanmurky
3
truly matter
3
caught
0
dropped / false+

honest toy: 50 standardized features, n = 64 samples; a real lasso (coordinate descent with soft-thresholding) and forward selection, computed live from a seeded random phenomenon. Nothing here is faked.

explained variance vs. how many causes you allow (k) — the Occam curve
3
one storythe whole spreadsheet
3
the elbow
0.00
R² at your k
0.00
signal ceiling
was it fair to drop the rest?
Move the k slider.
What you're watching
The Occam curve is explained variance as you're allowed more causes, one at a time, best first. On a genuinely sparse phenomenon it lunges up and then flattens: three features get you to the plateau, the next forty add almost nothing. That plateau is the sparse truth. The dashed line is the signal ceiling — the most any model could explain, the rest being irreducible noise. The gap between your k and the ceiling is real signal you left on the table.
02 · what to try

What to try

  1. Leave How many truly matter at 3. Watch the lasso paint three jade bars over the three ghosts and flatten everything else. This is clean recovery — essentialism vindicated: the phenomenon really did have a short cause.
  2. Drag Lasso strictness λ up. Watch marginal causes wink out first, the strongest surviving longest — the razor deciding, in order, what it can bear to drop.
  3. Find the elbow on the Occam curve, then set your k there. Note how little the ceiling rises past it. Adding causes has become vanity.
  4. Now drag How many truly matter up toward 40. The curve straightens, the elbow marches right, and the lasso starts dropping coral bars — real causes discarded to keep the story short. Hold k at 3 and read the verdict: the clean little story is now a lie.
03 · the honest core

Essentialism, rehabilitated and bounded

"What really matters here" has a bad reputation, and deservedly — it is the preferred sentence of the ideologue, the reductionist, the man with one idea. But strip the metaphysics and there is an empirical claim underneath, and it is often true: for a great many phenomena, a short list of causes carries nearly all the variance and the long tail carries nearly none. When that is so, the lasso recovers the short list without mysticism, and the elbow appears exactly where the causes run out. This is not a comforting story imposed on the data; it is a structure read off the data, testable and falsifiable.

So essentialism is rehabilitated — as a hypothesis. And it is bounded — because its validity is a bet about the phenomenon, not a law of thought. The sparse method does not assume the world is sparse; it discovers whether it is, and the residual keeps the score. The honest essentialist is the one who says "I think three things explain this" and then looks at what the three things leave out. The dishonest one skips the second clause.

04 · the discipline and its failure

Prefer few causes — then check the residual

Sparse thinking is a razor: prefer the explanation with fewer moving parts. But a razor with no stop cuts to the bone. The discipline has two clauses, and the second is the one people skip: prefer few causes, but check what a simple story leaves unexplained. If a three-cause story leaves half the variance in the residual, the world was dense and you are comforting yourself — the tidiness is yours, not the phenomenon's. The signal ceiling in the instrument is the honest referee: it tells you how much was ever explainable, so a large gap between your story and the ceiling is not humility, it is evasion.

The stakes are not only epistemic. "One root cause" is the grammar of every monocausal politics — the single enemy, the single reform, the single number that explains a society. Some problems really are sparse and yield to one lever. Many of the ones we argue about hardest — a personality, a civil war, why a generation is unhappy — are genuinely dense: dozens of small real causes, none dominant, none droppable without loss. Forcing a sparse story onto a dense world is not insight. It is the essentialist error wearing insight's clothes, and the residual is the tell.

05 · the mapping back

The same fit, different worlds

Once you hold the dial you see it everywhere. The good diagnostician and the conspiracy theorist run the same regression; they differ in whether they check the residual. The minimalist designer and the reductionist ideologue both bet on sparsity; one bets where the world obliges and one bets where it doesn't. The whole art is knowing which world you are in — and the only way to know is to fit the short story and then measure, honestly, what it failed to cover.

MathematicsLife
the featuresall the things that could matter — the whole list of candidate causes
the weight vectorhow much each one actually contributes
sparsitythe empirical fact that most of them contribute nothing
the recovered supportthe few that carry the phenomenon — the honest short list
the elbowthe diminishing returns of adding more causes
the residualwhat a simple story leaves out — your check on whether it was true

Most of what could matter, doesn't. Insight is finding the few that do — and the residual is how you find out whether you were kidding yourself.

06 · where the metaphor tears

Three honest rips

Sparsity lives in a basis
Sparsity is a property of the phenomenon and of the features you chose to describe it. The same reality can be sparse in one basis and dense in another: rotate the axes and three fat causes become thirty small ones, or the reverse. The "few causes" you recover may be an artifact of how you carved the variables — a different carving of the world would hand you a different short list, or no short list at all.
Many human phenomena are dense
Society, personality, history — these are the standing examples of genuinely dense phenomena, many small real causes with no dominant few. Forcing a sparse story onto them is exactly the essentialist error the word "sparsity" should warn you against. Drag the dial to 40 and watch: the method still returns a short, clean-looking answer. Its confidence has nothing to do with whether the world was sparse.
Correlated causes make the "essential" one arbitrary
When two candidate causes move together, the lasso keeps one and zeros its twin — not because the twin does nothing, but because either alone tells the same story. The recovered "essential cause" is then partly arbitrary: run it again, or nudge the data, and it may crown the sibling instead. Sparse methods can be confidently, cleanly, reproducibly wrong about which few matter.