The ASTRES Toolbox for Mode Extraction of

Non-Stationary Multicomponent Signals

Dominique Fourer? , Jinane Harmouche, Jeremy Schmitt, Thomas Oberlin,
Sylvain Meignen, Francois Auger and Patrick Flandrin
?
dominique@fourer.fr

AbstractIn this paper, we introduce the ASTRES toolbox TFRs Datadriven methods
which offers a set of Matlab functions for non-stationary multi-
component signal processing. The main purposes of this proposal STFT CWT Stransform EMD SSA
is to offer efficient tools for analysis, synthesis and transformation
of any signal made of physically meaningful components (e.g.
sinusoid, trend or noise). The proposed techniques contain some reassignment synchrosqueezing
recent and new contributions, which are now unified and theoret-
ically strengthened. They can provide efficient time-frequency or
time-scale representations and they allow elementary components ridge detection Mode extraction
extraction. Usage and description of each method are then
detailed and numerically illustrated.
Fig. 1. ASTRES toolbox content description freely available online2 .

I. I NTRODUCTION
Real world signals are often non-stationary and made of of the considered transforms and their TFRs are presented. In
several components. They require advanced techniques to be Section III, reassigned and synchrosqueezed versions of each
efficiently processed. Unfortunately, the Short-Time Fourier transform are described with two ridge detection methods. Sec-
Transform (STFT) and the Continuous Wavelet Transform tion IV describes data-driven methods, EMD and SSA, with
(CWT) which belong among the most usual approaches, are their corresponding new developments. Finally, the proposed
limited due to a poor energy localization in the time-frequency methods are illustrated by numerical experiments on real world
plane [1]. signal in Section V, before concluding the paper in Section VI.
Two solutions, the reassignment and the synchrosqueez- II. T IME - FREQUENCY REPRESENTATIONS
ing methods [2], [3], [4], can improve the readability of a A. Short-time Fourier transform (STFT)
time-frequency representation (TFR). More particularly, the Let Fxh (t, ) denote the STFT of signal x, using a differen-
synchrosqueezing can provide sharpened and invertible TFRs tiable analysis window h, defined as
allowing many applications like noise removal and signal Z
decomposition into modes [3], [5], [6]. In addition, the Fxh (t, ) = x(u)h(t u) eju du, (1)
R
Empirical Mode Decomposition (EMD) [7] and the Singular
Spectrum Analysis (SSA) [8], which belong to data-driven z being the complex conjugate of z. Thus, the spectrogram
methods, allow unsupervised signal decomposition into a sum can be computed by |Fxh (t, )|2 (cf. tfrstft, tfrgab). Eq.
of physically meaningful components (e.g. periodic functions, (1) admits the following synthesis formula when h(t0 ) 6= 0,
trend or noise). Several recent developments were proposed to allowing to recover signal x with a delay t0 0 (cf.
enhance these techniques [9] and to enable a fully automatic rectfrgab)
Z +
version of SSA, in the same flavor as EMD. 1 d
x(t t0 ) = Fxh (t, ) ej(tt0 ) . (2)
This paper proposes a collection of Matlab functions related h(t0 ) 2
to the new methods developed in the Analysis, Synthesis, An efficient recursive implementation (suitable for real-time
Transformation by Reassignment, EMD and Synchrosqueez- computation), is proposed in [10], [12]. It uses a specific
ing (ASTRES) research project . This new toolbox (cf. Fig. 1), causal (one-sided) analysis window which can be expressed
designed for automatic signal mode extraction, can be viewed k1
as hk (t)= T kt(k1)! et/T U (t), with k 1, where U (t) is the
as an extension of the previously proposed Time-Frequency
Heaviside step function and T a time spread parameter (cf.
ToolBox (TFTB)1 . It contains implementations of some recent
recursive_stft and stft_rec).
contributions related to reassignment, synchrosqueezing [10],
[11], [6], [12], EMD [13], [14], and SSA [9] methods. This B. Continuous wavelet transform (CWT)
paper is organized as follows. In Section II, proper definitions The CWT of a signal x is defined for an admissible mother
wavelet function as [1]
This research was supported by the French ANR ASTRES project (ANR- Z +  
13-BS03-0002-01). 1 t
1 http://tftb.nongnu.org, 2 https://github.com/dfourer/ASTRES toolbox Wx (t, s) = p x( ) d (3)
|s| s
If we define the scale as s = 0 , 0 being an arbitrary A. Reassignment
frequency, Eq. (3) can now be expressed as a time-frequency For each transform, the TFR values are moved according

transform as to the map (t, ) 7 (t(t,), (t,)), where t(t,) = Re t(t,)
s 
|| +
Z 
and (t,) = Im ( (t,)) (resp. s(t,) = Im(s(t,))), are the
CWx (t, ) = x( ) ( t) d. (4) reassignment operators.
0 0
1) STFT: Reassignment operators can be computed as [2]
The scalogram can be computed by |Wx (t, s)|2 or
|CWx (t, )|2 . If we use the Morlet wavelet [1], expressed as FxT h (t,) F Dh (t,)
t(t,) = t h
, (t,) = j + xh , (12)
1/4

t2
j0 t Fx (t,) Fx (t,)
(t) = T
e 2T 2 e , we finally obtain MWx (t, ) =
s with T h(t) = th(t) and Dh(t) = dh dt (t). Thus, the reas-
+ 2 (t )2
||
Z
2( T )2 signed spectrogram can be computed as (cf. tfrrsp and
x( ) e 0 ej( t) d (5)
0 T recursive_rsp) RFx (t, ) =
ZZ
(cf. tfrscalo and MW). The synthesis formula of Eq. (3) is
|Fxh (, )|2 t t(, ) ( (, )) d d. (13)


given by (cf. recMW) R2
Z Z
1 d A recent extension of the reassignment process, based on the
x(t) = Wx (t, s)|s|3/2 ds , with C = F ()
C R R Levenberg-Marquardt algorithm, allows to adjust the energy
(6) localization in the time-frequency plane through a damping
where F () denotes the Fourier transform of (t). parameter [17]. The new reassignment operators can be
C. Stockwell transform (S-transform) computed as
   
The S-transform [15] can be defined for a given zero-mean t (t, ) t
1 h
= t Rxh (t, ) + I2 Rx (t, ) (14)
signal x as (cf. tfrst) (t, )
 
Z + 2 (t )2 t t(t,)
 h 
|| with Rxh (t,) = , tRxh (t,) = R Rxh
STx (t, ) = x( ) e 2(0 T )2 ej d (7) (t,) t
x
(t,) (t,)
20 T
where T is the width of the Gaussian analysis window when where I2 is the 2 2 identity matrix. As a result, the
= 0 . The corresponding TFR, called Stockwellogram, is Levenberg-Marquardt reassigned spectrogram can be obtained
provided by |STx (t, )|2 .The S-transform can be viewed as a by replacing (t, ) by (t , ) in Eq. (13) (cf. tfrlmrgab
STFT using a frequency-varying window width, and recursive_lmrsp).
2) CWT: Since, t
Wx (t, s) = 1s WxD (t, s), the CWT
|| +
Z  

h
STx (t, ) = x( )h (t ) ej d (8) reassignment operators can be expressed by [2], [1], [18]
0 0
2 WxT (t, s)
where h is a Gaussian window defined as h(t) = e . 1
t
2T 2 t(t, s) = t + s , (15)
2T Wx (t, s)
The S-transform can also be related to the Morlet wavelet 0 s0 Wx (t, s)
transform by s(t, s) = = . (16)
s (t, s) WxD (t, s)
||
STx (t, ) = ejt MWx (t, ), (9) Then, the reassigned scalogram can be computed as (cf. rMW)
2 0 T
RWx (t, s) =
which shows that the S-transform can be viewed as a normal-
s02
ZZ
|Wx (, s0 )|2 t t(,s0 ) (s s(,s0 )) d ds0 .


ized phase-shifted Morlet wavelet transform. Thus, substitut-
ing MWx in Eq. (6) leads to the S-transform synthesis formula R2 s(,s0 )2
(17)
(cf. rectfrst2) [16]
Z 3) S-transform: Reassignment operators can be computed
1 d as for the STFT (cf. proofs in 8, [16]). Thus, both classical
x(t) = STx (t, ) ejt , (10)
Ch (0 T ) R || and Levenberg-Marquardt reassigned Stockwellograms are ob-
where Ch (0 T ) is a proportionality factor defined by tained through Eqs. (12) and (13), by replacing Fxh (t, ) by
Z SThx (t, ), FxT h by STTx h and FxDh by STDh x (cf. tfrrst.m
d and tfrlmrst). Thus we have RSTx (t, ) =
Ch (0 T ) = Fh ( 0 ) . (11)
R || ZZ
2

III. R EASSIGNMENT AND SYNCHROSQUEEZING |STx (, )| t t(, ) ( (, )) d d (18)
R2
Reassignment and synchrosqueezing are sharpening tech-
niques designed to improve TFRs. The main difference be- B. Synchrosqueezing
tween these methods is the reconstruction capability of syn- Each synchrosqueezed transform can be deduced from the
chrosqueezing and a quite poorer time-frequency localization simplified synthesis formula given by Eqs (2), (6) and (10),
compared to the reassignment, which is not invertible. respectively for the STFT, the CWT and the S-transform.
 1) STFT: The synchrosqueezed STFT can be defined For amplitude modulated signals, it is shown in [12] that
for any t0 0 such that h(t0 ) 6= 0 (cf. tfrsgab, (27) is biased, and can be improved using a slightly different
recursive_sstft) estimator expressed as (when |qx (t,)| < )
(3) (t, ) = Im (t,) + qx (t,)(t t(t,)) .
Z

0 (28)
h
SFx (t,) = Fxh (t, 0 ) ej (tt0 ) ( (t,0 )) d 0 , (19)
R An estimation of qx (t, ) can be computed for each transform
Its squared modulus provides a sharpened TFR, and it can be as follows:
inverted by Z 1) STFT: In [12], we proposed several local modulation
1 d estimators, valid for any differentiable analysis window, n
x(t t0 ) =
SFhx (t, ) , (20)
h(t0 ) R 2 2, which can be used in Eq. (27) or in (28) to obtain
Replacing by computed by Eq. (14) allows to make a vertical synchrosqueezed STFT transform (cf. tfrvsgab,
synchrosqueezing adjustable as for the Levenberg-Marquardt recursive_vstft)  n n1

reassignment (cf. tfrlmsgab, recursive_lmsstft). Re FxD h (t, )FxD h (t, )
2) CWT: The synchrosqueezed CWT can be defined as [3] xKn (t,) =
, (29)
Im FxT Dn1 h (t, )FxDn1 h (t, )
(cf. sMW) n n1
FxD h
FxD h (t,)FxDh (t,)
h
Z
(t,)Fx (t,)
SWx (t, s) = Wx (t, s0 )|s0 |3/2 (s s(t,s0 )) ds0 , (21) qx(tn) (t,) = Dn1 h n1 h ,
R Fx (t,)FxT h (t,) FxT D (t,)Fxh (t,)

and can be inverted by (cf. recsMW) (30)

n1 n2 n1

1
Z (FxT Dh
+ (n1)FxT h
)Fxh FxT h
FxDh
x(t) = SWx (t, s) ds, (22) qx(n) (t,) = n1 hF T h F T nhF h
,
C R FxT x x x
(31)
3) S-transform: The synchrosqueezed S-transform can be
where (t,) was omitted in Eq. (31) for the sake of clarity.
defined as [16] (cf. tfrsst, tfrlmsst)
In [12], these estimators were implemented and compared in
d 0
Z
0
SSTx (t, ) = || STx (t, 0 ) ej t ( (t, 0 )) 0 , (23) terms of accuracy for n = 2.
R | | 2) CWT: Using the local modulation estimator qx (t, s) =

t (t,s)
and can be inverted by (cf. rectfrsst) t proposed for the CWT [18], we obtain (cf. vsMW)
t (t,s)
Z
1 d
 
1 D2 D 2
x(t) = SSTx (t, ) , (24) s2 Wx (t,s)Wx (t,s) Wx (t,s)
Ch (0 T ) R || qx (t, s) = D . (32)
Wx (t,s)WxT (t,s) WxT D (t,s)Wx (t,s)
C. Second-order synchrosqueezing
3) S-transform: A new local modulation estimator can
Second-order synchrosqueezing was first proposed by Ober- also be derived to compute the vertical synchrosqueezed S-
lin et al. in [11] to improve TFRs of strongly modulated chirps transform [16] (cf. tfrvsst),
signal. Assuming a signal model expressed as 2
STD
h
x
Dh
(t,)STx (t,) STx (t,)
2
t 2 qx (t, ) = Dh Th T Dh
. (33)
x(t) = a(t) ej(t) , with (t) = x + x t + x , (25) STx (t,)STx (t,) STx (t,)STx (t,)
2
D. Ridge detection for mode extraction
where a(t) and (t) stand for the time-varying amplitude
and phase. Second-order synchrosqueezing proposes to use PLet us consider a multicomponent signal denoted by x(t) =
I ji (t)
local modulation estimation to improve the resulting TFR. i=1 xi (t). Then, each mode xi (t) = ai (t) e can be
Oblique synchrosqueezing (only proposed for the STFT) (cf. extracted from its synchrosqueezed transform, by restricting
tfrosgab) can be viewed as a time-frequency reassignment the integration area to the vicinity of the ridge, denoted i (t),
with a phase correction term [11], in each reconstruction formula given by Eqs. (20), (22) and
ZZ (24). The ridge i (t) can be directly estimated from a TFR
x using a ridge detector as:
OSFx (t, ) = Fxh (, 0 ) ej( 2 (t ))(t ) 1) Brevdo et al. method: This technique [19] aims at
R2
(t t(, 0 ))( (, 0 )) d d 0. (26) finding the best frequency curve (t) in the TFR Sx ,
which maximizes the energy with a smoothness constraint
Vertical synchrosqueezing uses an improved estimation of through a total variation penalization term expressed as (cf.
the instantaneous frequency (instead of the classical reassign- ridge_detect_brvmask)
ment operator), thanks to the chirp rate estimation x (t,) = Z Z 
 d 2

2
Im(qx (t,)) (see [12] for a further investigation). The improved = argmax |Sx (t, (t))| dt  (t) dt, (34)
frequency estimator can be computed as R dt R
( where controls the importance of the smoothness constraint.
(2) (t,) + x (t,)(t t(t,)) if |x (t,)| < For multi-component extraction (i.e. I > 1), this method can
(t, ) =
(t, ) otherwise, be iterated after setting Sx to zero in the vicinity of the
(27) previously detected ridge.
 2) Delaunay triangulation method: Frandrin proposes in To compare two classes Ci and Cj containing several
[6] a new way to disentangle the different components in time series, the dissimilarity is defined as the minimal
the time-frequency plane through a simplified representation distance between two distinct time series of each class
provided by the Delaunay triangulation attached to the spec- |< x, y >|
trogram zeros. For a particular Gaussian analysis window D(Ci , Cj ) = min d(x, y), d(x, y) = 1 ,
t2
xCi ,yCj ||x|| ||y||
g(t) = 1/4 e 2 , the zeros zn = n + jtn can be
determined from the STFT , through Bargmann Thus, each extracted mode is obtained by summing the ele-
Z transform as: mentary components included into the resulting classes Ci .
|z|2
Fxg (t, ) = e 4 Fx (z), with Fx (z) = A(z, )x( )d,
R V. N UMERICAL SIMULATIONS
2 2
1/4 2 j z+ z4
and A(z, ) = e . (35) A. Time-frequency representations
It admits a Weierstrass-Hadamard factorization such as [6]: Fig. 2 compares the TFRs of the gravitational wave signal
  GW150914 [21], resulting from a binary black hole collision.
z 2
These TFRs are computed using respectively, the Gabor trans-
e zn + 2 ( zn ) .
Y z 1 z
Fx (z) 1 (36)
n=1
zn form (STFT using a Gaussian analysis window), the Morlet
Wavelet transform and the S-transform, with their respec-
Then, the Delaunay triangulation is performed over zeros zn tive first- and second-order synchrosqueezed versions. The
and the domains attached to signal components are related to Reconstruction Quality Factor (RQF) given by RQF(x, x) =
outlier edges (cf. spz_delaunay_dom4a). 
||x||2

10 ln10 ||xx|| 2 is indicated above each invertible TFR.
IV. DATA - DRIVEN METHODS
The ASTRES toolbox also includes several implementations B. Empirical mode decomposition
of the well known EMD method [7], which was reformulated For this experiment (cf. Fig. 3), we consider a real-world
and extended for two dimensional signals [13], [14]. Due multi-component audio signal recorded from a cello. This
to paper size limitation, we chose to focus on more recent signal is analyzed using the synchrosqueezed STFT combined
developments allowing an unsupervised usage of SSA. with the Brevdo et al. method described in Section III-D1,
the EMD method and the proposed SSA method [9]. Syn-
A. Singular spectrum analysis
chrosqueezing leads to better separation results compared to
SSA [8] can expand a signal in a sum of periodic com- other methods which are more likely to create interference.
ponents, trends and noise. It considers a finite length time VI. C ONCLUSION AND FUTURE WORKS
series (e.g. a sampled signal) s = {sn , n = 1 . . . N } and a
user parameter L. Its algorithm (cf. ssa) is described by: The ASTRES toolbox was introduced as a collection of
Matlab functions for processing non-stationary and multicom-
1) Construction from x of the trajectory Hankel matrix X
ponent signals. This toolbox which is freely available online,
of size L K, with K = N L + 1.
unifies into the same framework several recent techniques
2) Singular Value Decomposition (SVD) of X, to obtain
R
developed into the ASTRES project. Some capabilities of
X = U V T =
X
Xi , with Xi = i ui vi T , (37) these methods designed for efficient TFRs computation and
i=1
mode extraction, were also illustrated by application examples
on real world signals. Future works consist in theoretically
i being the singular values of X and ui (resp. vi ) the strengthening these tools, and in proposing new applications.
column vectors of U (resp. V ).
R EFERENCES
3) Elementary components reconstruction from each Xi by
anti-diagonal averaging through [1] P. Flandrin, Time-Frequency/Time-Scale analysis. Acad. Press, 1998.
[2] F. Auger and P. Flandrin, Improving the readability of time-frequency

Xn and time-scale representations by the reassignment method, IEEE
n1 (Xi )m,nm+1 for 1 n < L


Trans. Signal Process., vol. 43, no. 5, pp. 10681089, May 1995.

m=1
[3] I. Daubechies, J. Lu, and H.-T. Wu, Synchrosqueezed wavelet trans-


X L forms: An empirical mode decomposition-like tool, Applied and Com-
(i) 1 putational Harmonic Analysis, vol. 30, no. 2, pp. 243261, 2011.
xn = L (Xi )m,nm+1 for L n K [4] F. Auger, P. Flandrin, Y. Lin, S. McLaughlin, S. Meignen, T. Oberlin,
m=1


and H. Wu, TF reassignment and synchrosqueezing: An overview,
L
X IEEE Signal Process. Mag., vol. 30, no. 6, pp. 3241, Nov. 2013.
1

(Xi )m,nm+1 for K + 1 n N.


N n+1
[5] S. Meignen, T. Oberlin, and S. McLaughlin, A new algorithm for
m=nK+1 multicomponent signals analysis based on synchrosqueezing: with an
application to signal sampling and denoising, IEEE Trans. Signal
4) Unsupervised mode extraction through elementary com- Process., vol. 60, no. 11, pp. 57875798, Nov. 2012.
[6] P. Flandrin, Timefrequency filtering based on spectrogram zeros,
ponents grouping, based on Hierarchical Clustering (HC) IEEE Signal Process. Lett., vol. 22, no. 11, pp. 21372141, 2015.
[9] (cf. ssa_decomp). The main idea of this algorithm [7] N. E. Huang, Z. Shen, S. R. Long, M. C. Wu, H. H. Shih, Q. Zheng, N.-
[20] is to look for the two nearest classes and to merge C. Yen, C. C. Tung, and H. H. Liu, The empirical mode decomposition
and the Hilbert spectrum for nonlinear and non-stationary time series
them to build a new class. This operation is iterated analysis, Proc. of the Royal Society of London A: Math., Physical and
until the desired number of classes (i.e. I) is reached. Engineering Sciences, vol. 454, no. 1971, pp. 903995, 1998.
 Gabor spectrogram, T=0.0223 Morlet scalogram, f0T=1.11 Stockwellogram, f0T=1.11

0.03 0.03 0.03

normalized frequency

normalized frequency

normalized frequency
0.025 0.025 0.025

0.02 0.02 0.02

0.015 0.015 0.015

0.01 0.01 0.01

0.005 0.005 0.005

500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
time samples time samples time samples

squ. mod. synch. STFT, T=0.0223, RQF=66.32 dB squ. mod. synch. MW, f0T=1.11, RQF=26.80 dB squ. mod. synch. ST, f0T=1.11, RQF=35.37 dB

0.03 0.03 0.03

normalized frequency

normalized frequency

normalized frequency
0.025 0.025 0.025

0.02 0.02 0.02

0.015 0.015 0.015

0.01 0.01 0.01

0.005 0.005 0.005

500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
time samples time samples time samples

squ. mod. vertical synch. STFT T=0.0223, RQF=44.89 dB squ. mod. vertical synch. MW, f0T=1.11, RQF=26.80 dB squ. mod. vertical synch. ST, f0T=1.11, RQF=32.25 dB

0.03 0.03 0.03

normalized frequency

normalized frequency

normalized frequency
0.025 0.025 0.025

0.02 0.02 0.02

0.015 0.015 0.015

0.01 0.01 0.01

0.005 0.005 0.005

500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
time samples time samples time samples

Fig. 2. Comparison between the Gabor STFT, Morlet CWT and the Stockell transform applied on the Livingston GW150914 signal [21].

synchrosqueezing extracted components EMD extracted components SSA extracted components

frequency [Hz]

frequency [Hz]
frequency [Hz]

4000 4000 4000

2000 2000 2000

0 0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
frequency [Hz]

frequency [Hz]

frequency [Hz]
4000 4000 4000

2000 2000 2000

0 0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
frequency [Hz]

frequency [Hz]

frequency [Hz]

4000 4000 4000

2000 2000 2000

0 0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
frequency [Hz]

frequency [Hz]

frequency [Hz]

4000 4000 4000

2000 2000 2000

0 0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
time [s] time [s] time [s]

(a) synchrosqueezed STFT (b) EMD (c) SSA

Fig. 3. Comparison of the extracted modes resulting from a real-world audio signal, using synchrosqueezed STFT, EMD and the proposed SSA method.

[8] J. Elsner and A. Tsonis, Singular Spectrum Analysis, A New Tool in Trans. Image Process., vol. 23, no. 12, pp. 52335248, 2014.
Time Series Analysis. Plenum Press, 1996. [15] R. Stockwell, L. Mansinha, and R. Lowe, Localization of the complex
[9] J. Harmouche, D. Fourer, F. Auger, P. Flandrin, and P. Borgnat, One or spectrum: the S-transform, IEEE Trans. Signal Process., vol. 44, no. 4,
two components ? the singular spectrum analysis answers, in Structured pp. 9981001, Apr. 1996.
Low-Rank Approximation (SLRA2015), Grenoble, France, Jun. 2015. [16] D. Fourer, F. Auger, and J. Hu, Reassigning and synchrosqueezing the
[10] D. Fourer, F. Auger, and P. Flandrin, Recursive versions of Stockwell transform: Complementary proofs, Tech. Rep., Dec. 2015.
the Levenberg-Marquardt reassigned spectrogram and of the syn- [Online]. Available: http://hal.archives-ouvertes.fr/hal-01467244
chrosqueezed STFT, in Proc. IEEE ICASSP, Mar. 2016, pp. 48804884. [17] F. Auger, E. Chassande-Mottin, and P. Flandrin, Making reassignment
[11] T. Oberlin, S. Meignen, and V. Perrier, Second-order synchrosqueezing adjustable: the Levenberg-Marquardt approach, in Proc. IEEE ICASSP,
transform or invertible reassignment? Towards ideal time-frequency Mar. 2012, pp. 38893892.
representations, IEEE Trans. Signal Process., vol. 63, no. 5, pp. 1335 [18] T. Oberlin and S. Meignen, The second-order wavelet synchrosqueezing
1344, Mar. 2015. transform, in Proc. IEEE ICASSP, Mar. 2017, p. (to appear).
[12] D. Fourer, F. Auger, K. Czarnecki, S. Meignen, and P. Flandrin, Chirp [19] E. Brevdo, N. S. Fuckar, G. Thakur, and H.-T. Wu, The synchrosqueez-
rate and instantaneous frequency estimation: application to recursive ing algorithm: a robust analysis tool for signals with time-varying
vertical synchrosqueezing, IEEE Signal Process. Lett., 2017. spectrum, Arxiv preprint, 2011.
[13] T. Oberlin, S. Meignen, and V. Perrier, An alternative formulation [20] J. H. Ward, Hierarchical grouping to optimize an objective function,
for the empirical mode decomposition, IEEE Trans. Signal Process., JASA, vol. 58, pp. 236244, 1963.
vol. 60, no. 5, pp. 22362246, 2012. [21] P. Flandrin, A note on the time-frequency analysis of GW150914,
[14] J. Schmitt, N. Pustelnik, P. Borgnat, P. Flandrin, and L. Condat, 2D Ecole normale superieure de Lyon, Tech. Rep., Sep. 2016. [Online].
prony-huang transform: A new tool for 2D spectral analysis, IEEE Available: https://hal-ens-lyon.archives-ouvertes.fr/ensl-01370441