“Chirps” everywhere

Patrick Flandrin*
CNRS — École Normale Supérieure de Lyon

*thanks to
Pierre-Olivier Amblard (LIS Grenoble), François Auger (Univ. Nantes),
Pierre Borgnat (ENS Lyon), Eric Chassande-Mottin (Obs. Nice),
Franz Hlawatsch (TU Wien), Paulo Gonçalvès (INRIAlpes),
Olivier Michel (Univ. Nice) and Jeffrey C. O’Neill (iConverse)
observation
 Doppler effect
Motion of a monochromatic source ⇒ differential perception of
the emitted frequency ⇒ “chirp”.

f+∆f f- ∆f "chirp"
 Pendulum

θ̈(t) + (g/L) θ(t) = 0

Fixed length L = L0
— Small oscillations are sinusoı̈dal, with

fixed period T0 = 2π L0/g.

“Slowly” varying length L = L(t) — Small oscillations
are quasi-

sinusoı̈dal, with time-varying pseudo-period T (t) ∼ 2π L(t)/g.
 Gravitational waves

Theory — Though predicted by general relativity, gravitational

waves have never been observed directly. They are “space-time
vibrations,” resulting from the acceleration of moving masses
⇒ most promising sources in astrophysics (e.g., coalescence of
binary neutrons stars).

Experiments — Several large instruments (VIRGO project for

France and Italy, LIGO project for the USA) are currently under
construction for a direct terrestrial evidence via laser interferom-
etry.
gravitational wave

time
VIRGO
 Bat echolocation

System — Active system for navigation, “natural sonar”.

Signals — Ultrasonic acoustic waves, transient (some ms) and

“wide band” (some tens of kHz between 40 and 100kHz).

Performance — Nearly optimal, with adaptation of emitted wave-

forms to multiple tasks (detection, estimation, classification, in-
terference rejection,. . . ).
 bat echolocation call + echo

time
bat echolocation call (heterodyned)

time
 More examples

Waves and vibrations — Bird songs, music (“glissando”), speech,

geophysics (“whistling atmospherics”, vibroseis), wide band pulses
propagating in a dispersive medium, radar, sonar,. . .

Biology and medicine — EEG (seizure), uterine EMG (contrac-

tions),. . .

Desorder and critical phenomena — Coherent structures in tur-

bulence, accumulation of precursors in earthquakes, “speculative
bubbles” prior a financial krach,. . .

Mathematics — Riemann and Weierstrass functions, . . .

description
 Chirps
Definition — We will call “chirp” any complex signal of the form
x(t) = a(t) exp{iϕ(t)}, where a(t) ≥ 0 is a low-pass amplitude
whose evolution is slow as compared to the oscillations of the
phase ϕ(t).

Slow evolution? — Usual heuristic conditions assume that:

1. |ȧ(t)/a(t)|  |ϕ̇(t)| : the amplitude is quasi-constant at the

scale of one pseudo-period T (t) = 2π/|ϕ̇(t)|.

2. |ϕ̈(t)|/ϕ̇2(t)  1 : the pseudo-period T (t) is itself slowly

varying from one oscillation to the next.
 Chirp spectrum
Stationary phase — In the case where the phase derivative ϕ̇(t)
is monotonic, one can approximate the chirp spectrum
 +∞
X(f ) = a(t) ei(ϕ(t)−2πf t) dt
−∞
by its stationary phase approximation X̃(f ). We get this way:
a 2 (t )
s
|X̃(f )|2 ∝ ,
|ϕ̈(ts)|
with ts such that ϕ̇(ts) = 2πf .

Interpretation — The “instantaneous frequency” curve ϕ̇(t) de-

fines a one-to-one correspondence between one time and one
frequency. The chirp spectrum follows by weighting the visited
frequencies by the corresponding times of occupancy.
 Signal in time

0.1

Real part
0

-0.1

Linear scale RSP, Lh=15, Nf=128, log. scale, Threshold=0.05%

0.45

0.4
Energy spectral density

0.35
Frequency [Hz]

0.3

0.25

0.2

0.15

0.1

0.05

0
20 10 0 50 100 150 200 250
Time [s]
representation
 Time-frequency

Idea — Give a mathematical formulation to musical notation

Objective — Write the “musical score” of a signal

Constraint — Get a localized representation in chirp cases:

ρ(t, f ) ∼ a2(t) δ (f − ϕ̇(t)/2π ) .

 Local methods and localization

The example of the short-time Fourier transform — One defines

the local quantity:
 ∞
x(s) h(s − t) e−i2πf s ds.
(h)
Fx (t, f ) =
−∞

Measure — Such a representation results from an interaction

between the analyzed signal and some apparatus (the window
h(t)).

Adaptation — Analysis adapted to impulses if h(t) → δ(t) and to

spectral lines if h(t) → 1 ⇒ adapting analysis to chirps requires
h(t) to be (locally) dependent on the signal.
 Self-adaptation of local methods
Matched filtering — If the window h(t) is the time-reversed signal
(x )
x−(t) := x(−t), one gets Fx − (t, f ) = Wx(t/2, f /2)/2, where
 +∞
Wx(t, f ) := x(t + τ /2) x(t − τ /2) e−i2πf τ dτ,
−∞
is the so-called Wigner-Ville Distribution (Wigner, ’32; Ville,
’48).

Linear chirps — The WVD localizes perfectly on straight lines

in the TF plane:
x(t) = exp{i2π(f0t + αt2/2)} ⇒ Wx(t, f ) = δ (f − (f0 + αt)) .

Remark — Localization via self-adaptation ends up in a quadratic

transformation (energy distribution).
 Beyond linear chirps

Global approach — The principle of self-adaptation via phase

compensation can be extended to non linear chirps (Bertrand &
Bertrand, ’84 ; Gonçalvès et F., ’94).

Limitations — Specific models and heavy computational burden.

Local approach — Spectrogram/scalogram = smoothed WVD

⇒ localized distributions via reassignment towards local centroı̈ds
(Kodera et al., ’76 ; Auger & F., ’94).
spectro

window = 21 63 127 points

signal model

Wigner-Ville
128 points
reass. spectro

window = 21 63 127 points

manipulation
 Chirps and dispersion

Example — Acoustic backscattering of an ultrasonic wave on

a thin spherical shell ⇒ frequency dispersion of elastic surface
waves.

acoustic backscaterring

time
 acoustic backscaterring
frequency

time
frequency

time
 dispersion compression
frequency

time frequency time

 Pulse compression

Limitation — Correlation radius ∼ 1/spectral bandwidth, ∀ signal

duration.

“Reception” — Post-processing by matched filtering (radar, sonar,

vibroseismics, non destructive evaluation).

“Emission” — Pre-processing by dispersive grating (production

of ultra-short laser pulses).
 Chirps and detection/estimation

Optimality — Matched filtering, maximum likelihood, contrast,. . . :

basic ingredient = correlation “observation — template”.

Time-frequency interpretation — Unitarity of a time-frequency

distribution ρx(t, f ) guarantees the equivalence:

| x, y
|2 = ρx, ρy

Chirps — Unitarity + localization ⇒ detection/estimation via

path integration in the plane.
 Time-frequency detection?
Language — Time-frequency offers a natural language for deal-
ing with detection/estimation problems beyond nominal situa-
tions.

Robustness — Uncertainties in a chirp model can be incorporated

by replacing the integration curve by a domain (example of post-
newtonian approximations in the case of gravitational waves).

gravitational wave
frequency

?
time
 Doppler tolerance

Signal design — Specification of performance by a geometrical

interpretation of the time-frequency structure of a chirp.

linear chirp hyperbolic chirp

frequency

frequency

time time
 bat echolocation calls (+ echo)
frequency

time
frequency

time
modeling
 Chirps and “atomic” decompositions

Fourier — The usual Fourier Transform (FT) can be formally

written as (F x)(f ) := x, ef
, with ef (t) := exp{i2πf t}, so that:
 +∞
x(t) = x, ef
ef (t) df.
−∞

Extensions — Replace complex exponentials by chirps, consid-

ered as warped versions of monochromatic waves, or by “chirplets”
(chirps of short duration) ⇒ modified short-time FTs or wavelet
transforms modifiées.
 Modified TFs — Example
Mellin Transform — A Mellin Transform (MT) of a signal x(t) ∈
L2(IR+, t−2α+1dt) can be defined as the projection:
 +∞
(Mx)(s) := x(t) t−i2πs−α dt = x, c
.
0

• Analysis on hyperbolic chirps c(t) := t−α exp{i2πs log t}.

• ϕ̇c(t)/2π = s/t ⇒, the Mellin parameter s can be interpreted

as a hyperbolic chirp rate.

• The MT can also be seen as a form of warped FT, since

x̃(t) := e(1−α)t x(et) ⇒ (Mx)(s) = (F x̃)(s).
 “Chirplets”
From “gaborets” and “wavelets” to “chirplets” — Localization
+ modulation lead to 4-parameter representations such as, e.g.,
x, xt,f,α,γ
with

xt,f,α,γ (s) ∝ exp{−π(γ + iα)(s − t)2 + i2πf (s − t)}.

Decomposition as estimation — Constitutive chirplets can be se-

quentially identified by “matching (or basis) pursuit” techniques
(Mallat & Zhang, ’93; Chen & Donoho, ’99; Bultan, ’99; Gribon-
val, ’99). They can also be estimated in the maximum likelihood
sense (O’Neill & F., ’98–’00).

“Parametric” limitation — Necessary trade-off between dictio-

nary size and algorithmic complexity.
“Chirplet” decomposition — An example

signal + noise 8 atoms

 Chirps and self-similarity

Dilation — Given H, λ > 0, let DH,λ be the operator acting on

processes {X(t), t > 0} as (DH,λX)(t) := λ−H X(λt).

Self-similarity — A process {X(t), t > 0} is said to be self-similar

of parameter H (or “H-ss”) if, for any λ > 0,
d
{(DH,λX)(t), t > 0} = {X(t), t > 0}.

Self-similarity and stationarity — Self-similar processes and sta-

tionary processes can be put in a one-to-one correspondence
(Lamperti, ’62).
 Lamperti

Definition — Given H > 0, the Lamperti transformation LH acts

on {Y (t), t ∈ IR} as:

(LH Y )(t) := tH Y (log t), t > 0,

and its inverse L−1
H acts on {X(t), t > 0} as :

(L−1
H X)(t) := e−Ht X(et), t ∈ IR.

Theorem — If {Y (t), t ∈ IR} is stationary, its Lamperti transform

{(LH Y )(t), t > 0} is H-ss. Conversely, if {X(t), t > 0} is H-ss, its
(inverse) Lamperti transform {(L−1 H X)(t), t ∈ IR} is stationary.
 tone

↑
Lamperti
↓
chirp
 “Spectral” representations

Fourier — (Harmonisable) stationary processes admit a spectral

representation based on Fourier modes (monochromatic waves):
 +∞
Y (t) = ei2πf t dξ(f ).
−∞

Mellin — (Multiplicatively harmonisable) self-similar processes

admit a corresponding representation based on Mellin modes
(hyperbolic chirps):
 +∞
X(t) = tH+i2πf dξ(f ).
−∞
 Weierstrass functions as an example

Discrete invariance — For any i.i.d. phases ϕn ∈ U (0, 2π) and

any 2π-periodic g, functions of the form
∞

WH,λ(t) = λ−Hn g(λnt) eiϕn ,
n=−∞
have a discrete scale invariance for any fixed λ, i.e., they are
H-ss only for µ = λk , k ∈ ZZ.

Consequence —Such “(H, λ)-DSI” processes have cyclostation-

ary Lamperti images and they can be represented on a discrete
basis of Mellin chirps (Borgnat et al., ’01).
 Weierstrass function (H = 0.5)

"delampertized" Weierstrass function

 Weierstrass-Mandelbrot

Fourier — In the case where g(t) = 1 − exp it, one gets the so-
called Weierstrass-Mandelbrot function, whose usual representa-
tion is given by a superposition of Fourier modes (in geometrical
progression).

Mellin — An equivalent representation exists (Berry and Lewis,

’80), as superposition of Mellin modes, i.e., of hyperbolic chirps.
 Weierstrass function (λ = 1.07; H = 0.3; tmax = 1; N = 1000; ν = 1)

original

time
detrended

time
 detrended Weierstrass function
frequency

time
 Chirps and power laws

A general model — Cα,β (t) = a tα exp{i(b tβ + c)}.

Example — Newtonian approximation of the inspiraling part of

gravitational waves → (α, β) = (−1/4, 5/8).

Typology — At t = 0: divergence of amplitude if α < 0, of

“instantaneous frequency” if β < 1 and of phase if β < 0.

Oscillating singularities. The case (α > 0, β < 0) is beyond a

simple Hölder characterization ⇒ development of specific tools
(2-microlocal analysis, wavelets).
 The Riemann function as an example
∞
Definition — σ(t) := n=1 n−2 sin πn2t

Differentiability — σ(t) happens to be non-differentiable if t = 

t0 = (2p + 1)/(2q + 1), p, q ∈ IN (Hardy, ’16) but differentiable in
t = t0 (Gerver, ’70).

Local chirps — One can show (Meyer, ’96) that, in the vicinity
of z = 1, the holomorphic version of Riemann function reads
∞

σ(1 + z) = σ(1) − πz/2 + Kn(z) C3/2,−1(z),
n=1

leading to σ(1 + t) = σ(1) − πt/2 + O(|t|3/2) when t → 0.

 Riemann function

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

frequency

0.955 0.96 0.965 0.97

time
conclusion
 Chirps and time-frequency

Signals — Chirps “everywhere”

Representations — Natural description framework = the time-

frequency plane

Models — “Chirps = time-frequency trajectories” ⇒ the notion

of instantaneous frequency can be approached as a by-product
of representations in the plane (e.g., “ridges”, or fixed points of
reassignment operators)