the second hundred · metaphor 208
Can you truly know someone from only a few glimpses? Sometimes yes — if what matters about them is simple, a handful of the right moments is enough to rebuild the rest. The whole catch lives in that word, simple.
We reconstruct people constantly from fragments. A few unguarded seconds — how he treats a waiter, what he does when he thinks no one's watching, the one story he can't stop telling — and we feel we have the shape of him. Most of the time we're not being lazy. If a person's truth is concentrated in a few strong traits, then a few well-aimed observations really can recover almost all of it, and the rest follows.
Mathematics makes this exact and, at first, unbelievable. If a signal is sparse — mostly zero, with only a few spikes that carry it — you can recover it perfectly from far fewer measurements than its length, provided the measurements are the right, scattered kind. Not a summary. The whole thing, exactly. The field is called compressed sensing, and it turns "a few glances were enough" from a hope into a theorem — with a sharp edge marking exactly when it stops being true.
The surprising idea
The old rule of thumb says you need at least as many measurements as unknowns: to pin down a signal of length 64, take 64 readings. Compressed sensing breaks that rule by charging for a different thing. It says: what costs you is not the length of the signal but its complexity — the number of pieces that are actually nonzero, its sparsity k. A signal that is 64 numbers long but only k=4 of them nonzero is, in the way that matters, a four-number thing wearing a sixty-four-number coat.
The trick has two parts. First, the measurements must be incoherent — each one a random scatter across the whole signal, so no glance duplicates another. Second, among all the signals consistent with your handful of readings, you pick the simplest one: the sparsest, found by minimizing the sum of absolute values (L1) or by a greedy hunt that grabs the strongest-matching spike, then the next (matching pursuit). Astonishingly, when the true signal is sparse enough, that simplest explanation is the true one — not close, exact — from roughly k·log N measurements, not N.
What to try
Start where it lands: a few spikes, a handful of measurements, teal sitting exactly on violet — error near zero, verdict recovered. Now slide measurements down. For a while nothing breaks; you are stealing glances and still rebuilding the whole. Then you cross the frontier and the recovery shatters — teal spikes appear where the truth is silent, the error jumps, the verdict flips to failed. That cliff is the whole story: it is not gradual.
Now put measurements back and raise sparsity instead — make the truth complicated, spread across many components. Watch the same collapse from the other side: the more of the person there is to know, the more glances it takes, until no small handful suffices and you'd need to see nearly everything. The map below the signal shows this as a single sweep — a green territory of the recoverable and a red one of the lost, split by a clean diagonal. You live near that line. Press New signal to redraw the truth and confirm the frontier stays put; it depends on k and m, not on luck.
The mapping
We are compressed sensors of each other, and mostly it works. If what's essential about someone is concentrated — a few load-bearing values, a couple of deep fears, one way they always break — then a few well-chosen encounters carry enough to reconstruct the rest, and our confidence is earned, not arrogant. The measurements that work are the incoherent ones: the offhand moment, the stranger's account, the behavior under stress — glances that each catch a different facet rather than the same rehearsed surface twice.
But the theorem also names its own limits, and they are the honest ones. The recovery is exact only if the truth really is sparse. A person whose meaning is spread thin across ten thousand small contingencies — no dominant traits, everything mattering a little — cannot be recovered from a handful; you would need to witness nearly all of them, and a first impression will be confidently, precisely wrong. And the glances must be the scattering kind: see someone only in the one context they've mastered, and you have taken the same measurement over and over, learning almost nothing new.
Read as life lessons
It isn't the amount of you that sets the price of being understood — it's the sparsity. Whoever's truth is concentrated in a few strong things can be grasped from a few moments. Complexity is what demands the long acquaintance.
Measurements only count if they're incoherent — each catching something new. Ten looks at the same polished surface teach less than three from scattered angles. Variety of vantage, not volume of contact.
Below the frontier, recovery is exact; a hair above it, nonsense. Understanding doesn't fade smoothly with too few glimpses — it holds, then shatters. A confident wrong read and a true one can sit one measurement apart.
In the wild
An MR image is sparse in the right basis, so scanners now sample a fraction of the data and reconstruct the full picture — turning a long scan a child can't hold still for into a short one.
The "single-pixel camera" and countless sensor arrays take far fewer readings than pixels, betting the scene is sparse in wavelets — capturing more with less hardware and power.
Pooled genetic testing and wideband spectrum sensing lean on the same fact: when only a few things are "on," a few clever combined measurements find them all.
The mapping, exactly
| Mathematics | Life |
|---|---|
| the sparse signal | A person or thing whose truth is concentrated in a few strong features, most of it near-blank. |
| sparsity k | How complicated they really are — the count of things that genuinely carry who they are. |
| a measurement | One glimpse — an encounter, a story, a moment observed — that mixes many facets into a single reading. |
| incoherence | Seeing them from scattered, independent angles, so no two glances repeat the same information. |
| L1 / matching pursuit | Choosing the simplest explanation consistent with what you saw — the fewest assumptions that fit. |
| the recovery frontier | The sharp line between "a handful was enough" and "you never really had them." |
The honest model
There is no cheating here. The instrument builds a genuine length-N=64 signal x with exactly k random nonzero spikes, and a fixed measurement matrix A whose entries are random Gaussians — the incoherent, scattering measurements. It computes the true readings y = A·x, keeping only m of them, then throws the signal away and tries to rebuild it from y and A alone, using orthogonal matching pursuit: repeatedly seize the component most correlated with what's left unexplained, re-fit by least squares, and subtract. The recovered x̂ is compared to the truth; the recovery error is the real relative distance ‖x̂−x‖ / ‖x‖.
The map runs that whole procedure hundreds of times — once per square, sweeping sparsity against measurement count — and colors each by whether the error fell below a strict threshold. Nothing is scripted to a nice picture; the diagonal frontier emerges, and it lands where the theory says it must, near m ≈ 2k·log(N/k). "Toggle a little noise" adds real Gaussian corruption to the measurements, and you can watch exact recovery soften into approximate — the honest fate of the idealization in a noisy world.
Where the metaphor tears
The whole miracle rests on a precondition that human beings routinely violate: that the truth is simple. A life whose meaning is smeared across thousands of small contingencies — no dominant traits, context all the way down — has no sparse core to recover, and a confident read from a handful of glimpses isn't insight, it's a projection. The math tells you exactly when you're entitled to generalize, and the honest answer is often "not yet."
The theorem needs measurements that scatter across the whole and repeat nothing. Real acquaintance is the opposite: we see people in the same few settings, at the same times, playing the same roles — coherent, redundant sampling that piles up readings without adding information. You can watch someone for years through one keyhole and recover almost nothing.
Perfect reconstruction assumes clean measurements and a truly sparse signal. Add noise — and every real observation is noisy — and "exact" degrades to "approximately," with a fidelity that depends on how far past the frontier you are. Toggle the noise on and watch: the guarantee doesn't vanish, but it stops being a promise and becomes a bet.