Tikhonov Regularization
Imagine you are reconstructing a medical image from noisy scanner data, or recovering a temperature
history from a smudged sensor. You solve the equations exactly — and out comes a jagged,
wildly-oscillating mess that no doctor or engineer would trust. The exact fit chased the
noise, not the signal. This is the curse of an
ill-posed problem:
insisting on a perfect fit amplifies tiny measurement errors into enormous, meaningless swings.
Tikhonov regularization is the single most important cure in all of inverse
problems. Its idea is almost embarrassingly simple, and profound: don't fit the data
perfectly. Instead, balance two goals at once — fit the data
reasonably well, and keep the solution small, smooth, and sensible.
A single knob, the parameter \alpha, dials between them.
The bargain, written as maths
Naïve inversion trusts the data completely, even where the operator is blind. Tikhonov strikes a
bargain: fit the data and keep the model small. It minimises a sum of two terms,
\hat m_\alpha = \arg\min_m \Big( \|Gm - d\|^2 + \alpha^2\,\|m\|^2 \Big).
The first term is the
data misfit —
how badly the model m reproduces the data d.
The second is a penalty on the size of the model, weighted by the
regularization parameter \alpha. A small
\alpha barely penalises anything → a closer but noisier fit; a large
\alpha punishes any large model hard → a smoother but blurrier answer.
(Exactly how to set that knob is its own art — that comes up next in
choosing the regularization parameter.)
Setting the gradient to zero gives a modified normal equation:
(G^{\mathsf T}G + \alpha^2 I)\,\hat m_\alpha = G^{\mathsf T}d.
Adding \alpha^2 I lifts every eigenvalue of
G^{\mathsf T}G away from zero — a matrix that was singular (and so had
no unique inverse) becomes not just invertible but well-conditioned. This
modified system is always solvable, no matter how nasty the original problem. And it is
exactly
ridge regression
from machine learning, here cast as the cure for ill-posedness.
Worked example 1 — one number, and why the penalty saves you
Strip everything down to a single unknown. Suppose G = g is just a
number, so the data is d = g\,m + \text{noise}. The naïve inverse is
m = d/g. If the operator is nearly blind in this direction —
g tiny, say g = 0.01 — then a noise wobble of
0.1 in d becomes a
0.1/0.01 = 10 swing in m: a hundred-fold
blow-up. The Tikhonov solution instead is
\hat m_\alpha = \frac{g\,d}{g^2 + \alpha^2}.
With g = 0.01 and a modest \alpha = 0.1, the
denominator is dominated by \alpha^2 = 0.01, not by
g^2 = 0.0001. That noise-driven direction is gently switched off instead
of exploding. In a strong direction (g = 5,
\alpha = 0.1) the denominator is essentially
g^2 and you recover d/g almost untouched.
Strong directions kept, weak directions tamed — from one tiny algebraic tweak.
Worked example 2 — what it does to each singular value
The one-number story generalises perfectly through the
SVD.
Tikhonov multiplies each reconstructed component by a filter factor:
\hat m_\alpha = \sum_i f_i\,\frac{u_i^{\mathsf T}d}{\sigma_i}\,v_i, \qquad f_i = \frac{\sigma_i^2}{\sigma_i^2 + \alpha^2}.
Look at f_i. Where \sigma_i \gg \alpha
(strong, trustworthy directions) f_i \approx 1 — kept intact. Where
\sigma_i \ll \alpha (weak, noise-prone directions)
f_i \approx \sigma_i^2/\alpha^2 \approx 0 — smoothly switched off. The
dangerous 1/\sigma_i blow-up is tamed: instead of dividing by a tiny
number, the filter sends that term toward zero.
Try it: with \alpha = 0.1, a big singular value
\sigma = 2 gives f = 4/(4 + 0.01) = 0.9975
(kept), while a small one \sigma = 0.02 gives
f = 0.0004/(0.0004 + 0.01) = 0.039 (crushed). The crossover, where
f_i = \tfrac12, sits exactly at \sigma_i = \alpha.
Worked example 3 — a gentle dimmer, not a sharp switch
Compare Tikhonov with its blunter cousin,
truncated SVD (TSVD).
TSVD uses a hard filter — it keeps a direction fully
(f_i = 1) if \sigma_i is above a cutoff and
throws it away completely (f_i = 0) if it is below. Tikhonov's filter
\sigma_i^2/(\sigma_i^2+\alpha^2) is a soft ramp that
slides smoothly from 1 to 0 as \sigma_i shrinks past
\alpha.
Think of TSVD as a light switch — on or off — and Tikhonov as a dimmer. The dimmer
avoids the abrupt "ringing" artefacts that a hard cutoff can leave behind, because it doesn't slam a
near-cutoff direction to zero all at once. Both damp the noisy tail of the spectrum; Tikhonov just
does it continuously.
The filter in action
The curve is the filter factor f_i across singular directions (high
index = small \sigma_i). Turn up \alpha and
the cutoff marches to lower indices — more aggressive smoothing, suppressing more directions. Turn
it down toward zero and every factor approaches 1, recovering the unstable naïve inverse.
Choosing \alpha is choosing where to draw that line.
- Minimise \|Gm-d\|^2 + \alpha^2\|m\|^2; solve (G^{\mathsf T}G + \alpha^2 I)m = G^{\mathsf T}d.
- Filter factors f_i = \sigma_i^2/(\sigma_i^2 + \alpha^2): ≈1 for strong directions, ≈0 for weak ones.
- It is ridge regression, and it trades data fit against model size via \alpha.
Here is the trap that surprises everyone: the Tikhonov solution is not the exact
least-squares answer, and it is not meant to be. By penalising the model's size you pull
it systematically toward zero — you build in a small, deliberate bias. Why on
earth would you want a biased answer?
Because the exact answer, for an ill-posed problem, has enormous variance: it
lurches all over the place as the noise wiggles. Tikhonov trades a tiny, controlled amount of
systematic error for a huge reduction in noise sensitivity. This is the famous
bias–variance trade-off, and it is why there is no single "perfect"
\alpha — only a best compromise for your noise level. Too small
and variance floods back; too large and bias smears the real answer away. Finding that balance is
exactly the job of
choosing the regularization parameter.
This exact bargain was discovered three separate times, in three fields that barely spoke
to each other:
- In applied mathematics and geophysics, the Russian mathematician
Andrey Tikhonov developed it (partly for
Soviet geophysical inverse problems) — and it bears his name.
- In statistics, the very same penalty appeared as ridge
regression, the standard fix for unstable regression coefficients.
- In neural networks, it turns up yet again as weight decay —
shrinking the network's weights toward zero to stop it over-fitting.
Add a penalty on the size of the solution, and instability and over-fitting both melt away. Three
communities, working independently, converged on the identical mathematics — one of the great
cases of convergent evolution in mathematics.