The Picard Condition
Before you pour weeks into reconstructing a hidden image, a subsurface structure, or a deblurred
photograph, wouldn't you love a quick test that says "this data can be sensibly inverted"
— or "give up, there is no stable answer hiding in here"? That is exactly what the
discrete Picard condition gives you: a single, honest go/no-go diagnostic for an
ill-posed inverse problem, run before you waste any effort.
The trick is to look at the data through the eyes of the
SVD.
The reconstruction is
\hat m = \sum_i \dfrac{u_i^{\mathsf T} d}{\sigma_i}\, v_i, a sum of
terms, each one a data coefficient |u_i^{\mathsf T} d| divided by a
singular value \sigma_i. Since the singular values march downhill toward
zero, the only way each ratio stays small — the only way the sum converges to something sensible —
is if the data coefficients decay faster than the singular values. That is the
Picard condition, in one sentence.
When it holds, every ratio |u_i^{\mathsf T} d|/\sigma_i stays under
control and the reconstruction is meaningful. When it fails — as noise always eventually forces it
to — the ratios blow up and the answer is meaningless static.
The signature of noise
Why should a real signal satisfy Picard while noise wrecks it? Because the two live in completely
different parts of the spectrum. Real, physical signals are smooth: their content sits in
the first few singular directions, so the high-index coefficients
|u_i^{\mathsf T} d| fall away quickly — often even faster than the
singular values. Noise has no such courtesy. It is roughly white: it spreads about
equally across all the singular directions, so the noise contribution
|u_i^{\mathsf T} e| does not decay at all — it flattens out at a
noise floor while \sigma_i keeps plunging.
The result is a tug of war with a clear winner. For small i the true
signal dominates the coefficient and everything is fine. Past the index where the flat noise floor
overtakes the shrinking singular values, the ratio
|u_i^{\mathsf T} d|/\sigma_i turns upward and diverges. That crossover
index is precious: it is exactly where to stop trusting the data — it tells you how many singular
components are worth keeping, and so sets the
truncation level
and the strength of regularization.
Worked example: reading the crossover on a log scale
Here is the practical recipe. Compute the SVD, then plot two sequences together against the
singular index i, always on a logarithmic
y-axis (the numbers span many orders of magnitude): the singular values
\sigma_i, and the data coefficients
|u_i^{\mathsf T} d|. Take a concrete toy spectrum where
\sigma_i = e^{-i} and the noise floor sits at
10^{-3}:
\begin{array}{c|c|c|c}
i & \sigma_i = e^{-i} & |u_i^{\mathsf T} d| & \text{ratio} \\ \hline
1 & 0.368 & 0.368 & 1.0 \\
3 & 0.050 & 0.050 & 1.0 \\
5 & 0.0067 & 0.0067 & 1.0 \\
7 & 0.0009 & 0.001\ (\text{floor}) & 1.1 \\
9 & 0.00012 & 0.001\ (\text{floor}) & 8.1
\end{array}
Read the table left to right. While the data coefficient tracks the singular value (rows
i = 1 to 5), the ratio holds steady at about
1 — Picard is satisfied and those components are gold. Around
i = 7 the coefficient stops falling and pins to the noise floor of
10^{-3}, while \sigma_i keeps shrinking. From
there the ratio climbs — 1.1, then 8.1, and
it will keep exploding. The crossover near i \approx 7 is your cutoff:
keep the first six or seven components, throw the rest away.
Worked example: what "levelling off" really means
The single most important visual cue on a Picard plot is the moment the data coefficients
level off. Think about what is happening physically. Each coefficient
u_i^{\mathsf T} d = u_i^{\mathsf T}(\text{signal}) + u_i^{\mathsf T}(\text{noise})
is a signal part plus a noise part. Early on, the signal part is much bigger, so the coefficient
rides down with the true signal. Later, the signal part has decayed below the noise part, so the
coefficient can fall no further — it flattens at
|u_i^{\mathsf T} e| \approx the noise level.
Beyond that flattening point, dividing by an ever-tinier \sigma_i does
nothing but amplify noise. So the flattening index is not a nuisance — it is a gift. It
draws a bright line on the plot: everything to the left is recoverable signal, everything to the
right is noise wearing a signal's clothes. That line is precisely where to truncate. Move the noise
slider in the figure below and watch the flattening point — and hence the safe cutoff — slide left
as the noise floor rises.
The Picard plot
Three quantities versus singular index: the singular values \sigma_i
(decaying), the data coefficients |u_i^{\mathsf T} d| (decaying until
they hit the noise floor), and their ratio. With no noise the ratio stays flat — Picard satisfied.
Add noise and the coefficients flatten; past the crossover the ratio shoots up, marking the useful
components from the hopeless ones.
- The solution is well-behaved when |u_i^{\mathsf T}d| decays faster than \sigma_i.
- Noise flattens the data coefficients at a floor; beyond the crossover, |u_i^{\mathsf T}d|/\sigma_i diverges.
- The crossover index sets how many components to keep — the natural regularization cutoff.
It is tempting to hope that with careful enough measurement your data might satisfy Picard all the
way out. It won't. Real noisy data always eventually violates the Picard condition.
No matter how clean your instrument, beyond some index the coefficients
|u_i^{\mathsf T} d| flatten out at the noise level and stop decaying,
while the singular values keep marching to zero. From that point on, every extra term you include
divides a fixed lump of noise by a smaller and smaller \sigma_i.
So blindly summing all the SVD components — which is exactly what the plain
pseudoinverse
does — is not "using all your information"; it is a guaranteed noise-dominated disaster. The Picard
plot is not a warning you might ignore; it shows you precisely where the useful data ends and where
you must stop.
The condition is named after the French mathematician Émile Picard, whose early-1900s work on
integral equations asked, in the continuous world of infinitely many singular values,
when a Fredholm equation of the first kind actually has a square-integrable solution. His answer —
the generalized Fourier coefficients of the data must decay fast enough relative to the singular
values — was a piece of pure functional analysis, about as far from an engineering worksheet as you
could imagine.
A century later that same abstract criterion, discretized, has become one of the most
practical tools in the inverse-problems toolbox. Practitioners routinely compute the SVD
and glance at the "Picard plot" the way a mechanic glances at a dashboard gauge: is the ratio flat
(good), and where does it turn up (there is my cutoff)? It is a lovely example of deep theory
quietly graduating into an everyday, visual go/no-go check.