Priors Are Regularization
Up to now, regularization has looked like a slightly embarrassing trick: your least-squares fit
blew up, so you added a penalty to calm it down, and you tuned a mysterious knob
\alpha until the answer looked sensible. Where does that penalty
come from? Why \|m\|^2 and not something else? And what on earth
is \alpha?
The Bayesian view answers all three at once, and the answer is startlingly clean:
Tikhonov regularization is exactly what you get when you assume the solution has a
Gaussian prior probability distribution and the noise is Gaussian, and then ask
for the most probable solution given the data. The penalty is not a trick bolted on
afterwards — it is the negative logarithm of a prior belief. And the mysterious
\alpha turns out to be a precise, interpretable ratio: how noisy you
think the data is, divided by how large you expect the model to be.
Here is the punchline that unites the two halves of this course. Take the Bayesian
MAP objective
with a zero-mean Gaussian prior m \sim N(0, C_M) and uniform Gaussian
noise C_D = \sigma^2 I. The negative log-posterior is, up to constants,
-\log p(m\mid d) \;=\; \frac{1}{\sigma^2}\|d - Gm\|^2 \;+\; m^{\mathsf T} C_M^{-1} m.
That is exactly a
general Tikhonov
functional. The data-fit term is the likelihood; the penalty term is the prior. Maximising the
posterior is regularized least squares.
Where the equivalence comes from
Bayes' theorem says the posterior — what you believe about
m after seeing the data — is proportional to the
likelihood times the prior:
p(m \mid d) \;\propto\; p(d \mid m)\, p(m).
Now feed in Gaussians. Gaussian measurement noise means the likelihood is
p(d\mid m) \propto \exp\!\Big(-\tfrac{1}{2\sigma^2}\|d - Gm\|^2\Big),
and a zero-mean Gaussian prior on the model is
p(m) \propto \exp\!\Big(-\tfrac{1}{2\tau^2}\|m\|^2\Big).
Because these are exponentials of quadratics, taking -\log of the
product turns the multiplication into a sum and strips the exponentials away, leaving a plain sum
of squared norms. The MAP estimate — the model that maximises the
posterior — is therefore the model that minimises
\frac{1}{2\sigma^2}\|d - Gm\|^2 + \frac{1}{2\tau^2}\|m\|^2,
which, after multiplying through by 2\sigma^2, is the Tikhonov
functional \|d-Gm\|^2 + \alpha^2\|m\|^2 with
\alpha^2 = \sigma^2/\tau^2. Maximising a Gaussian posterior and
minimising a Tikhonov objective are the same act in two languages.
The dictionary
Term by term, the Bayesian and deterministic pictures are the same object in two languages:
- A prior N(0, \tau^2 I) ⇒ the penalty \|m\|^2/\tau^2 — standard Tikhonov with \alpha^2 = \sigma^2/\tau^2.
- A smoothness prior (covariance favouring smooth fields) ⇒ a derivative penalty \|Lm\|^2 — general Tikhonov, with L^{\mathsf T}L = C_M^{-1}.
- A confident prior (small \tau) ⇒ a large \alpha — heavy regularization.
So the regularization parameter was never arbitrary: \alpha is the
ratio of how noisy you think the data is to how large you expect the model to be. And the penalty
was never a mere mathematical trick — it is a precise statement of prior belief. The
negative log-prior is the penalty, and the parabola
m^2/(2\tau^2) is the L2 penalty drawn out.
Worked examples: reading the knobs
1. The meaning of \alpha
Because \alpha^2 = \sigma^2_{\text{noise}}/\sigma^2_{\text{prior}}, the
regularization strength is a tug-of-war between two variances. Suppose your sensors are
noisy, \sigma^2_{\text{noise}} = 4, and you expect a fairly large model,
\sigma^2_{\text{prior}} = 100. Then
\alpha^2 = 4/100 = 0.04 — light regularization, because you mostly trust
that the model can be large. Now quieten the sensors to
\sigma^2_{\text{noise}} = 0.01: \alpha^2 = 0.0001,
almost no regularization, because clean data should be trusted. More noise ⇒ more
regularization; a tighter prior ⇒ more regularization. Both are exactly what intuition
demands, and now they fall out of a formula rather than a hunch.
2. A tighter prior pulls harder
Hold the noise fixed at \sigma^2_{\text{noise}} = 1. If you strongly
believe the model is small, set a tight prior \sigma^2_{\text{prior}} = 1,
giving \alpha^2 = 1. Loosen your belief to
\sigma^2_{\text{prior}} = 25 and
\alpha^2 = 0.04 — the penalty weakens twenty-fivefold and the fit is
allowed to roam. The width of your prior is the strength of your penalty, inverted.
3. A smoothness penalty is a correlated prior
The general penalty \|Lm\|^2 from
smoothing Tikhonov
corresponds to a prior with covariance C_M = (L^{\mathsf T}L)^{-1}. A
diagonal L = I means the model components are independent under
the prior — knowing one tells you nothing about its neighbour. But a derivative
L = D_2 makes C_M couple neighbouring values:
the prior now says "adjacent points are correlated, so they should look alike."
That correlation is exactly what makes the reconstruction smooth. A smoothness penalty and a
smoothness prior are the same statement.
The penalty drawn from the prior
The bell curve is a zero-mean Gaussian prior on a model component; the upward parabola is its
negative logarithm — the penalty m^2/(2\tau^2) that regularization
adds. Narrow the prior (smaller \tau, more confident) and the parabola
steepens — a stronger pull toward zero, i.e. a larger \alpha.
- Negative log-posterior = data misfit + (negative log-prior) = a Tikhonov functional.
- Gaussian prior N(0,\tau^2 I) ⇒ L2 penalty with \alpha^2 = \sigma^2/\tau^2; smoothness prior ⇒ derivative penalty.
- \alpha is the noise-to-prior ratio: regularization strength is a statement of belief.
Digging deeper
The Bayesian reading is illuminating, but it comes with a warning printed in the fine text. Your
regularized solution is only ever as good as your prior assumptions. The MAP
estimate is the most probable model under the prior you chose — and if the true solution
is nothing like a sample from your assumed Gaussian, that "most probable" answer can be
confidently, systematically wrong.
Suppose the truth is sparse — mostly zeros with a few big spikes — or has hard
edges, neither of which looks anything like a smooth Gaussian draw. A Gaussian prior will
spread each spike into a gentle bump and blur each edge into a ramp, and it will do so with total
statistical confidence. The optimisation succeeded perfectly; it just optimised the
wrong objective. No amount of clean data or careful tuning of
\alpha can rescue a bad modelling assumption. The math cannot
fix a wrong prior — it can only faithfully carry it through to a wrong answer.
It really is. Once you see regularization as "likelihood times Gaussian prior", an astonishing
range of tools collapse into one family, all running the same Bayesian machinery with Gaussian
priors under the hood:
- Ridge regression in statistics — literally Tikhonov with L = I.
- The Kalman filter — a Gaussian prior (the prediction) updated by a Gaussian likelihood (the measurement), step after step in time.
- Gaussian-process regression — a smoothness prior over functions, which is a correlated Gaussian prior exactly like the C_M = (L^{\mathsf T}L)^{-1} above.
And here is the jump that changes everything: swap the Gaussian prior for a sparsity
prior — a Laplace distribution, whose negative log is
|m| rather than m^2 — and the L2 penalty
becomes the L1 / LASSO penalty behind compressed sensing. That is the
technology that lets an MRI scanner reconstruct a sharp image from a fraction of the usual
measurements, imaging you in a fraction of the usual time. Different prior, different penalty,
same beautiful machine.
The whole equivalence rides on one small algebraic miracle. A Gaussian is
e^{-(\text{quadratic})}. Multiplying the likelihood by the prior
multiplies two such exponentials; taking -\log converts that product
into a sum of the quadratics and deletes the exponential entirely. So
"maximise a product of Gaussians" becomes "minimise a sum of squared norms" — and a sum of
squared norms is precisely a least-squares-plus-penalty objective. Gaussians are the one prior
family for which the negative log is a clean quadratic, which is exactly why they, and
not some other bell-ish curve, give you Tikhonov on the nose.