The Bayesian Formulation
This is the fullest, most powerful way to pose an inverse problem — and, remarkably, it fits in one
line. Bayes' theorem says
the posterior distribution of the unknown model given the data is proportional to
the likelihood (how well a model explains the data) times the prior
(what we believed about the model beforehand). That is the entire idea. Everything else on this page
is unpacking it.
p(m \mid d) \;\propto\; \underbrace{p(d \mid m)}_{\text{likelihood}}\; \underbrace{p(m)}_{\text{prior}}.
What makes this so much more than a formula is that three separate things you used to bolt
on by hand are now all present at once: regularization, uncertainty, and prior knowledge.
They are not add-ons; they are consequences. Where the
statistical view
told us the answer is a distribution, the Bayesian formulation tells us exactly how to
build that distribution.
The posterior is not a single model
The Bayesian view reframes the whole enterprise. The answer to an inverse problem is not a single
model but a probability distribution over models — the posterior
p(m \mid d), which captures both the best estimate and how sure we are
of it. The likelihood carries the physics and the noise; the prior
carries everything we knew before the experiment (smoothness, positivity, a rough magnitude).
Read the two factors as two voices. The likelihood p(d\mid m) is the
data speaking: "models that reproduce my measurements score high." A Gaussian noise model
makes this voice a least-squares data-misfit term. The prior p(m) is
prior expectation speaking: "models that are smooth / small / positive score high." A
Gaussian prior makes this voice a regularization term. The posterior is the two voices multiplied —
the models both voices like.
Gaussian everything
Take Gaussian noise e \sim N(0, C_D) and a Gaussian prior
m \sim N(m_{\text{prior}}, C_M). Both factors are Gaussian, so the
posterior is Gaussian too, and its peak — the
MAP estimate — minimises the
sum of two Mahalanobis terms:
\hat m = \arg\min_m\Big[(d - Gm)^{\mathsf T}C_D^{-1}(d - Gm) + (m - m_{\text{prior}})^{\mathsf T}C_M^{-1}(m - m_{\text{prior}})\Big],
\hat m = m_{\text{prior}} + \big(G^{\mathsf T}C_D^{-1}G + C_M^{-1}\big)^{-1}G^{\mathsf T}C_D^{-1}\,(d - G m_{\text{prior}}).
The prior term C_M^{-1} is added to G^{\mathsf T}C_D^{-1}G
— and that addition is exactly what makes the matrix invertible. The prior is the
regularizer, a point the next page makes precise.
Worked example — Tikhonov falls out of the negative log
Let's watch the famous Tikhonov functional appear on its own. Suppose the noise is
isotropic Gaussian, C_D = \sigma_D^2 I, and the prior is isotropic
Gaussian centred at zero, C_M = \sigma_M^2 I. Maximising the posterior is
the same as minimising its negative logarithm (the log turns the product into a sum,
and Gaussians have e^{-(\cdot)} shapes). Taking
-\log of p(d\mid m)\,p(m) gives
-\log p(m\mid d) \;=\; \frac{1}{2\sigma_D^2}\|Gm - d\|^2 \;+\; \frac{1}{2\sigma_M^2}\|m\|^2 \;+\; \text{const}.
Multiply through by 2\sigma_D^2 (which doesn't move the minimiser) and it
reads
\hat m = \arg\min_m\Big[\|Gm - d\|^2 + \lambda^2\|m\|^2\Big], \qquad \lambda^2 = \frac{\sigma_D^2}{\sigma_M^2}.
That is the familiar Tikhonov
functional, exactly — and it wasn't invented, it was derived. Even better, the mysterious
regularization parameter \lambda now has a meaning: it is the
ratio of the noise variance to the prior variance. Noisy data (large
\sigma_D) or a confident prior (small \sigma_M)
both push \lambda up, leaning harder on the prior. Clean data or a vague
prior push it down, leaning harder on the measurement. The dial was never arbitrary.
Prior and data combine
The one-parameter picture: a prior belief about a model component, the
likelihood the data provides, and the posterior that fuses them.
Tighten the prior (small C_M) to lean on prior knowledge; tighten the
data (small C_D) to lean on the measurement. Either way the posterior is
narrower than both — combining information always reduces uncertainty.
- The solution is the posterior p(m\mid d) \propto p(d\mid m)\,p(m).
- Gaussian likelihood + Gaussian prior ⇒ Gaussian posterior; its mean/mode is the MAP estimate.
- The prior precision C_M^{-1} added to G^{\mathsf T}C_D^{-1}G stabilises the inversion — the prior is the regularizer.
Bayes is honest, but it is not magic. The posterior is only as trustworthy as the two ingredients
you feed it: the prior and the noise model. Choose a smooth
Gaussian prior when the true solution is actually sparse and edgy — a piecewise-constant
image, a spiky spectrum — and the method will confidently smear those edges into gentle ramps.
Misjudge the noise level \sigma_D and you set the wrong
\lambda: too small and noise leaks through, too large and real features get
flattened.
The saving grace is that the framework is honest about what it assumes — every
assumption is written down as a distribution you can inspect and change. But it cannot rescue you
from a wrong assumption. So the prior is a genuine modelling choice, on a par with choosing
the physics, not a neutral technicality to be picked by default. Choosing it well is where the craft
lives.
Step back and the Bayesian formulation is the grand unifying picture of inverse
problems. It shows that regularization was never an arbitrary hack — it is a precise
statement of prior belief, spelled out in the open. It shows that the real "answer" is a
whole posterior distribution, from which you can extract not just a best estimate but full error
bars (the
posterior covariance).
And it shows that inverse problems are the same machinery humming inside a startling amount
of modern technology. The Kalman filter steering a rocket or a robot vacuum is a recursive Bayesian
update. The denoiser in your phone's camera is a posterior with an image prior. Medical scanners
reconstruct pictures the same way, and much of modern AI is Bayesian inference wearing a large hat.
Learn p(m\mid d) \propto p(d\mid m)\,p(m) once, and you have met the engine
of half the century.