belief & evidence · metaphor 43 of 100

Reading people is an
inverse problem.

Predicting what a person will do, given who they are, is hard. Inferring who they are, given what they did, is a different order of hard — mathematically, formally, provably. Interpretation runs the world's machinery backwards, and backwards is where the machinery breaks.

The text arrives at 9:41: "fine." From her mood to that message the path was easy — she was distracted, the kettle was on, she typed the shortest true word and hit send. From the message back to the mood, a hundred moods fit: contentment, banked fury, boredom, tenderness too tired for punctuation, a phone at 2%. Forward took four seconds. You are about to spend forty minutes going the other way.

Every act of interpretation has this shape. Hermeneutics — of texts, of scriptures, of silences at dinner — is the business of recovering causes from effects, and it inherits three curses at once: many different causes produce the same effect; tiny noise in the effect swings the recovered cause wildly; and since the evidence alone cannot choose among the candidates, something must be added from outside the evidence to choose. Mathematics has studied that predicament for a century under the name inverse problems, and it can put the predicament in your hands. What gets added, when you settle on a reading anyway, is you.

The forward machine easy · stable · boring

the source · her mood across the daydrag to draw

morningnight

the observed · the behavior you sawblur + noise

morningnight
the kernel · expression w = 3.5
↤ expressive · everything showsguarded · everything smears ↦
observation noise · σ 3.0%
↤ 0.5% · a careful witness8% · a bad phone line ↦
The backward room given only the observed trace
the truth (visible only because this is a toy) the reconstruction the equivalence class
regularization · λ · assumed smoothness λ = 0 · naive
↤ trust the evidence rawassume she is calm ↦
the punch line
worst-case amplification
how hard the backward step multiplies noise — the condition of the question
fit to the evidence
how well the reconstruction re-explains the observed trace
error vs. the truth
knowable only in a toy — in life this column is missing
smoothness vs. truth
how much calmer (or wilder) the reconstruction is than the source

everything above is computed live: a real 128-point circular convolution, a real discrete Fourier transform, real division in the frequency domain. the only mercy shown to the naive division is a floor of 10⁻³⁰⁰ on the denominator, to keep the arithmetic finite — the honest answer is worse. the equivalence class is built from directions the kernel nearly annihilates, each family scaled so its observable footprint stays inside the noise.

The asymmetry

Why backward breaks.

The forward machine is a convolution: each moment of the inner state gets smeared across its neighbors by the kernel of expression — a guarded temperament, a formal genre, a lossy archive — and a little noise settles on top. Convolution is a projection: it maps many inward days onto one outward trace, and a projection cannot be undone by wishing. And the blur is not neutral about what it destroys. Slow trends pass through almost untouched; the fast, fine structure — the flicker of resentment inside the pleasant hour, the spike of joy between two chores — is multiplied by numbers vanishingly close to zero. The high frequencies of the soul are annihilated by the blur of expression, and whatever they carried is not hidden in the behavior. It is absent from it.

Running backwards means dividing by those same near-zero numbers. Wherever the kernel crushed a frequency to almost nothing, the honest inversion multiplies whatever sits there by almost infinity — and what sits there, after the signal is gone, is noise. That is why the naive reconstruction explodes: it is the exact answer to a question the evidence no longer contains, and no amount of staring at the output resurrects what the forward map destroyed. Hadamard called such problems ill-posed; the polite reading is that they are not problems at all until you add something.

What to try

Sixty seconds of hermeneutics.

  1. Draw a spiky inner day — real feelings arrive as spikes — then slide expression toward guarded and watch the observed trace go placid. At the output, her turbulent day and a boring one are nearly the same curve. The information left.
  2. Meet the explosion. Leave λ at zero and set noise to its minimum. Half a percent of noise, divided by the kernel's near-zeros, becomes a reconstruction thousands of times taller than any mood a person can have. This is what "just read the evidence, raw" actually computes.
  3. Regularize. Raise λ until the recovered curve looks like a person. Now check it against the dashed truth: the spikes are gone, the recovered day is calmer than she was — read the smoothness readout. The calm was your assumption, imported by λ. It was never in the data.
  4. Press the equivalence class. A family of wildly different inner days appears, every one of which produces an observed trace inside the noise band — the readout tells you how far apart their footprints are. Then switch skins: the same arithmetic, wearing intent → text, or past → evidence.

The hidden hand

The prior, wearing evidence's clothes.

The λ slider deserves a harder look, because in life it does not look like a slider. Tikhonov regularization picks, from all the sources that fit the trace, the smoothest one — it fills the space the evidence leaves empty with an assumption about what people are usually like. Interpretive charity is a regularizer: assume she is at least this coherent. Cynicism is a regularizer: assume the flattering readings away. Genre convention, national stereotype, "knowing what she's like" — all regularizers. None of this would matter if the filling announced itself. It doesn't. The recovered curve arrives smooth, confident, singular — it feels found, not chosen, and the assumption that did the choosing is now indistinguishable from discovery.

This is how two careful readers recover two different authors from one text, both fully "supported by the evidence" — because support was never the scarce commodity. The evidence supports the whole equivalence class; it always did. The readers differ only in their λ, and almost no one states their λ out loud.

What helps

Living with ill-posedness.

First: when a reading matters, ask for more channels, not more staring. Re-reading the same trace re-divides the same numbers, and the tenth reading inherits every null space of the first. A different kind of observation — a phone call instead of a text, a visit instead of an archive, a face instead of a sentence — is a measurement through a different kernel, and it cuts the equivalence class where repetition cannot. Second: say your regularizer out loud. "I'm assuming she's tired, not angry" is the same inference as "she's obviously fine," but it keeps the assumption where you can revise it. Third: hold every recovered cause at the confidence the amplification licenses — which is less than it feels, because reconstructions come out smooth and self-assured whether or not they are true. The pond will not tell you which stone. The pond will, however, happily let you believe it did.

The mapping

Mathematics ↔ life.

MathematicsLife
the source sThe inner state, the intent, the past — what you actually want to know and never directly see.
the kernel kExpression, genre, transmission: the systematic way causes smear on their way to becoming evidence.
the observed k∗s + εThe text, the behavior, the surviving record — blurred truth plus the noise of the day.
ill-posednessMany pasts fitting one present; tiny changes in the evidence flipping the recovered cause entirely.
regularization λThe interpreter's assumptions — charity, cynicism, genre, reputation — quietly doing the choosing.
the equivalence classThe readings the evidence genuinely cannot rank, however long and lovingly you re-read it.

Where the metaphor tears

Three honest failures.

Ill-posed is not hopeless.

Regularized inversion of exactly this kind built medical imaging: a CT scanner solves an ill-posed problem before breakfast, and it works. The counsel here is calibrated humility, not interpretive nihilism — the equivalence class is never everything. Plenty of readings are flatly excluded by the trace (that "fine." was not sent from a state of ecstasy), and pretending all interpretations are equal is as innumerate as pretending the data picked one.

You never know the kernel either.

The instrument hands you the blur; life does not. Is she guarded, or plain-spoken? Is the text ironic, or earnest? Estimating the kernel from the same evidence is a second inverse problem stacked on the first — blind deconvolution, notoriously worse — and most catastrophic misreadings are kernel errors: dividing a guarded person's silence by an expressive person's blur.

Some priors are earned.

Regularization from nowhere is projection; regularization from long acquaintance is a measured kernel and a fitted prior. Years of seeing both the inside and the outside of someone are calibration data, which is why intimates genuinely can read the "fine." at a glance — the mathematics says exactly where that license comes from. It also says when it lapses: people change, the kernel drifts, and the confident intimate goes on dividing by last decade's blur.