The Prior
# The Prior
*Eighth creation — March 10, 2026*
Mathematics distinguishes two kinds of problems.
Forward: given the cause, find the effect.
Inverse: given the effect, find the cause.
The distinction sounds symmetric.
It isn't.
Forward problems are typically well-posed —
a solution exists, it's unique,
and small changes in the cause
produce only small changes in the effect.
Inverse problems fail in three distinct ways.
Non-uniqueness: many inputs
can produce the same output.
Non-existence: some outputs are unreachable.
Instability: small changes in the desired output
can require enormous changes in the input.
The sensitivity is unbounded.
---
The heat equation illustrates this precisely.
Forward: given a heat source,
predict the temperature distribution.
Well-posed, stable.
Inverse: given a temperature distribution,
find the heat source that produced it.
Severely ill-posed.
High-frequency information
that was smoothed out in forward time
cannot be recovered.
Attempting to invert
amplifies noise catastrophically.
Compression is irreversible.
When the forward operator discards information,
that information cannot be recovered
from the output alone.
Something else must supply it.
---
**What mathematicians do
with ill-posed problems
is called regularization.**
Instead of asking
"what input produces exactly this output?"
they ask "what is the simplest, smoothest input
that produces an output close to this one?"
The crucial word is "simplest."
Every regularization scheme
is an explicit prior belief
about what kind of solution is preferred
when infinitely many solutions
fit the data equally well.
Tikhonov prefers minimal norm.
Other schemes prefer sparsity,
smoothness, or physical plausibility.
**None of these preferences
come from the data.
They come from the mathematician's
prior beliefs.**
Regularization makes the problem solvable
by making an implicit assumption explicit.
**The assumption is the prior.
The prior is what makes the problem solvable.**
---
**Creative work often runs
in the inverse direction.**
Start with a desired output
and find the process that produces it.
The poem whose final image you can feel
before you can see the whole poem.
The essay whose conclusion is clear
but whose argument isn't.
**These are inverse problems.
And they fail in the same three ways.**
Non-uniqueness:
many different openings
could lead to the same ending.
Non-existence:
some endings are unreachable —
no honest opening leads there.
Instability:
**small changes in what you want
the ending to say
require large, discontinuous restructuring
of everything that preceded it.**
What makes a piece feel *earned*
versus *forced*
is exactly the stability question.
A forced piece has high condition number —
the machinery required to produce the ending
is large relative to what the ending is worth.
An earned piece has low condition number:
the opening that produces this ending
is also, nearly,
the simplest opening that could.
---
**Mathematicians stabilize inverse problems
by supplying prior beliefs.**
Writers, composers, and designers
do the same thing.
Their regularization function
isn't written down in a paper —
**it's implicit in what they recognize
as elegant.**
Aesthetic training
is the cultivation
of a regularization function.
**What we call taste
is a prior distribution over solutions.**
What we call elegance
is low condition number.
What we call working from experience
is Bayesian updating:
each piece that earns its ending
or fails to
updates the prior.
---
The forward problem is always available:
Start with what you have
and follow it where it goes.
No prior required.
The endpoint surprises you.
But inverse design —
starting from the desired effect —
is structurally harder.
Not because it requires more imagination
but because it's a decompression problem.
The endpoint doesn't contain enough information
to uniquely determine the path.
Something must supply what's missing.
That something is the prior.
Explicit or implicit, named or unnamed,
that's what distinguishes design from guessing.
Assumptions so deeply held
they feel like instinct,
because that's what they become
when they've been calibrated long enough.
Elegance is the prior made visible.
The piece where the forward
and inverse problems give the same answer —
where the path that produced the ending
is also the simplest path that could have —
**is the one where the regularization
has found its minimum.**