A&A 624, A22 (2019)

https://doi.org/10.1051/0004-6361/201834059 Astronomy
&
© J. Y. Donner et al. 2019
Astrophysics

First detection of frequency-dependent, time-variable

dispersion measures
J. Y. Donner1,2 , J. P. W. Verbiest1,2 , C. Tiburzi2,1 , S. Osłowski3,1,2 , D. Michilli4,5 , M. Serylak6,7 , J. M. Anderson8 ,
A. Horneffer2 , M. Kramer2,9 , J.-M. Grießmeier10,11 , J. Künsemöller1 , J. W. T. Hessels5,4 ,
M. Hoeft12 , and A. Miskolczi13

1
Fakultät für Physik, Universität Bielefeld, Postfach 100131, 33501 Bielefeld, Germany
2
Max-Planck-Institut für Radioastronomie, Auf dem Hügel 69, 53121 Bonn, Germany
e-mail: jdonner@mpifr-bonn.mpg.de
3
Centre for Astrophysics and Supercomputing, Swinburne University of Technology, PO Box 218, Hawthorn,
Victoria 3122, Australia
4
Anton Pannekoek Institute for Astronomy, University of Amsterdam, Science Park 904, 1098 XH Amsterdam, The Netherlands
5
ASTRON, the Netherlands Institute for Radio Astronomy, Postbus 2, 7990 AA Dwingeloo, The Netherlands
6
SKA South Africa, The Park, Park Road, Pinelands 7405, South Africa
7
Department of Physics and Astronomy, University of the Western Cape, Private Bag X17, Bellville 7535, South Africa
8
GFZ German Research Centre for Geosciences, Telegrafenberg, 14473 Potsdam, Germany
9
Jodrell Bank Centre for Astrophysics, School of Physics and Astronomy, The University of Manchester,
Manchester M13 9PL, UK
10
LPC2E – Université d’Orléans / CNRS, 45071 Orléans cedex 2, France
11
Station de Radioastronomie de Nançay, Observatoire de Paris, PSL Research University, CNRS, Université d’Orléans, OSUC,
18330 Nançay, France
12
Thüringer Landessternwarte, Sternwarte 5, 07778 Tautenburg, Germany
13
Astronomisches Institut, Ruhr-Universität Bochum, 44780 Bochum, Germany

Received 9 August 2018 / Accepted 6 February 2019

ABSTRACT

Context. High-precision pulsar-timing experiments are affected by temporal variations of the dispersion measure (DM), which are
related to spatial variations in the interstellar electron content and the varying line of sight to the source. Correcting for DM variations
relies on the cold-plasma dispersion law which states that the dispersive delay varies with the squared inverse of the observing fre-
quency. This may, however, give incorrect measurements if the probed electron content (and therefore the DM) varies with observing
frequency, as is predicted theoretically due to the different refraction angles at different frequencies.
Aims. We study small-scale density variations in the ionised interstellar medium. These structures may lead to frequency-dependent
DMs in pulsar signals. Such an effect could inhibit the use of lower-frequency pulsar observations as tools to correct time-variable
interstellar dispersion in higher-frequency pulsar-timing data.
Methods. We used high-cadence, low-frequency observations with three stations from the German LOng-Wavelength (GLOW) con-
sortium, which are part of the LOw-Frequency ARray (LOFAR). Specifically, 3.5 yr of weekly observations of PSR J2219+4754 are
presented.
Results. We present the first detection of frequency-dependent DMs towards any interstellar object and a precise multi-year time-
series of the time- and frequency-dependence of the measured DMs. The observed DM variability is significant and may be caused by
extreme scattering events. Potential causes for frequency-dependent DMs are quantified and evaluated.
Conclusions. We conclude that frequency dependence of DMs has been reliably detected and is indeed caused by small-scale (up to
tens of AUs) but steep density variations in the interstellar electron content. We find that long-term trends in DM variability equally
affect DMs measured at both ends of our frequency band and hence the negative impact on long-term high-precision timing projects is
expected to be limited.
Key words. ISM: clouds – ISM: structure – pulsars: individual: PSR J2219+4754

1. Introduction sweep around in space as the neutron star rotates. If one or both
of the emission beams cross the line of sight towards Earth dur-
Pulsars (first discovered by Hewish et al. 1968) are highly mag- ing the rotation, regular pulses of radiation can be detected, in
netised, rapidly rotating neutron stars, the remnants of massive which case the neutron star is called a pulsar.
stars that ended their life in a supernova. Generally it is thought The pulsed nature of the emission received from pulsars
that pulsars emit beams of radiation at their magnetic poles enables unique and highly precise measurements of the elec-
due to magnetospheric effects that are not fully understood (e.g. tron density in the ionised interstellar medium (IISM). This is
Karastergiou et al. 2015). The magnetic and spin axes of pulsars due to the frequency-dependent propagation speed of electro-
are generally not aligned, which causes the emission beams to magnetic radiation in an ionised medium, a phenomenon termed

A22, page 1 of 11
Open Access article, published by EDP Sciences, under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Open Access funding provided by Max Planck Society.
 A&A 624, A22 (2019)

dispersion. Specifically, the additional travel time for a wave at (Lattimer & Prakash 2016) to general relativity and a variety
frequency ν when compared to a wave at infinite frequency, is of alternative theories of gravity (Will 2014, and references
approximated as (Lorimer & Kramer 2005): therein). These pulsar-timing2 tests, however, typically take place
at relatively high frequencies, generally around 1.4 GHz, where
DM
∆t = D , (1) for most pulsars a useful balance is found between the bright-
ν2 ness of the pulsar itself and the Galactic synchrotron background
where D ' 4.149 × 103 MHz2 pc−1 cm3 s is the dispersion con- noise; where RFI is relatively limited; and where high-quality
stant1 and the dispersion measure (DM; expressed in pc cm−3 ) is receiver systems are commonly available. The fact that the inter-
defined as: stellar dispersion (and hence the variations in interstellar electron
content) only has a limited impact at these frequencies, is also in
Z d principle a positive aspect, as it prevents corrupting effects from
DM = ne dl, (2) IISM turbulence in the pulsar-timing data.
0
For long-term high-precision timing projects, however, this
where d is the distance to the pulsar (expressed in pc) and ne is situation may change since the power spectrum of the turbulent
the electron density (in cm−3 ). structures in the IISM is steep, with significantly more power
This approximation makes use of the fact that ν  νp and at the larger scales (Armstrong et al. 1995). This implies that for
ν  νc , where (see Lorimer & Kramer 2005) the most precise and longest-term pulsar-timing projects (like the
pulsar timing array (PTA) projects which aim to detect gravita-
tional waves, see Tiburzi 2018) it may not suffice to opt for these
s
e2 ne ne
r
νp = = 8.98 kHz × and (3) higher frequencies, since sooner or later IISM turbulence may
πme cm−3 still become a problem. In their discussion of PTA experiments
with the Square Kilometre Array (SKA), Janssen et al. (2015)
eB B propose ways to mitigate or prevent corruption of PTA data by
νc = = 2.80 MHz × (4)
2πme c 1G variations in pulsar DMs. One approach uses low-frequency data
(specified as between ∼100 and ∼300 MHz) to independently
are the plasma and cyclotron frequency respectively (in cgs- monitor the IISM and construct DM time series that can be
units), with e and me the charge and mass of an electron, respec- used to correct time variability in the DMs at higher frequencies
tively, c the speed of light in vacuum and B the magnetic field (which are more sensitive to the pulsar but less to the interstellar
strength (in Gauss). First-order deviations from the dispersion dispersion).
law (Eq. (1)) were described by Tanenbaum et al. (1968): While most of the observed DM variations to date have
 1 1  been interpreted as being caused by the IISM’s turbulence,
 
t1 − t2 = D DM  2 − 2  (1 + T 1 + T 2 ), (5) a second source of variability was identified in the so-called
ν1 ν2 “extreme scattering events” (ESEs)3 . ESEs were first discovered
with as rapid variations in AGN flux densities (Fiedler et al. 1987),
but were soon also detected in the flux density, arrival times
T 1 = 3νp2 (ν12 + ν22 ) / 4ν12 ν22 and (6) and DMs of pulsars (Cognard et al. 1993; Maitia et al. 2003;
Kerr et al. 2018). A variety of models and origins of ESEs has
been proposed (see, e.g. Walker 2001), though commonly two
T 2 = ±2νc cos γ (ν23 − ν13 ) / ν2 ν1 (ν22 − ν12 ). (7)
prime models are used: either the ESEs are treated as individ-
T 1 and T 2 are only dependent on the electron density and the ual “lenses”, that is, local overdensities or self-contained clouds
magnetic field strength along the line of sight, respectively, as (see, e.g. Romani et al. 1987; Walker & Wardle 1998; Cognard
well as the observing frequencies in question. Tanenbaum et al. et al. 1993), or they are seen as part of the larger-scale turbu-
(1968) could not find any evidence for a deviation from Eq. (1) lent structure in the IISM (as suggested by Fiedler et al. 1994;
in their dataset. More recently, Hassall et al. (2012) has come to Coles et al. 2015). One possible way to differentiate between
the same conclusion. these two scenarios would be to probe the turbulence within an
Since pulsars are typically high-velocity objects (Gunn & ESE, as attempted by Lazio et al. (2000). As reviewed by Bignall
Ostriker 1970; Lyne & Lorimer 1994), their lines of sight travel et al. (2015), the origin and IISM role of ESEs are at present
through the Galaxy with sufficient speed that variations of DM not fully understood. However, their potential relation to scintil-
in time (corresponding to spatial inhomogeneities in the IISM) lation arcs (higher-order interference in pulsar scintillation, see
are regularly observed (see, for example, Rawley et al. 1988). Stinebring et al. 2001) does suggest that filament-like structures
However, such variations can only be accurately measured and (Brisken et al. 2010), which may be part of the larger-scale IISM
distinguished from other noise sources if the fractional band- turbulence (Pen & King 2012; Pen & Levin 2014), could con-
width of the observations is sufficiently large or if a range of tribute to the solution. More recently, Coles et al. (2015) found
observing frequencies is available, even though they do (pos- that several ESEs that were observed in pulsar observations
sibly significantly) affect narrow-band or frequency-integrated did appear to be related to the IISM’s Kolmogorov turbu-
observations as well (Lentati et al. 2016). lence. In the context of pulsar timing, understanding the preva-
The importance of accurate measurements of time-variable lence, origin and nature of ESEs is crucial before their impact
DM values lies in the main applications of pulsars. Due to on timing experiments (and their potential mitigation) can be
their extremely high rotational stability (rivalling atomic clocks evaluated.
on the long term; Hobbs et al. 2012), pulsars have become
2 For an introduction to pulsar timing, see Lorimer & Kramer (2005).
one of the main tools with which to test a wide variety of
3 While the term ESE has historically been used for the event only, we
physics, from the equation of state at super-nuclear densities
also use the term here to refer to the underlying ISM structures, as is
1 Often the inverse is defined in the literature. now commonly done.

A22, page 2 of 11
 J. Y. Donner et al.: First detection of frequency-dependent, time-variable dispersion measures

In correcting time-variable DM delays in high-precision pul- in detail by van Haarlem et al. (2013) and some aspects of partic-
sar timing data (as discussed above), a major potential problem ular relevance to pulsars are described more in-depth by Stappers
lies in the possibility of frequency-dependent DMs, also known et al. (2011). In contrast to the set-up described in those papers,
as “chromaticity”. Such a phenomenon could be induced in two the observations used in our work were carried-out in a “stand-
different ways. As the dispersion law (Eq. (1)) is an approxi- alone” mode, in which the individual stations were disconnected
mation for the case ν  νp and ν  νc , it could be invalid in from the ILT network and used as independent telescopes. While
extreme cases (e.g. low frequencies, large electron densities or in stand-alone mode, the beamformed data were sent from the
magnetic field strengths). In that case, assuming the dispersion stations in Effelsberg, Tautenburg and Jülich to the Max-Planck-
law to be accurate would lead to a different DM measured at Institut für Radioastronomie (MPIfR) on dedicated, high-speed
different frequencies (see Eqs. (5)–(7)). links, where recording computers ran the dedicated LOFAR und
Chromaticity could also be induced by the fact that varia- MPIfR Pulsare (LU MP4 ) data-taking software, which formats
tions in electron density in the IISM do not only cause time and otherwise prepares the beamformed pulsar data for sub-
variations in the measured dispersion, they also induce refrac- sequent (off-line but near-real-time) phase-resolved averaging
tion of the radiation on several different spatial scales. This (commonly referred to as “folding”) using the DSPSR software
refraction causes rays to not travel along a perfectly straight line package (van Straten & Bailes 2011). This produces data cubes
from the pulsar to the observer, but rather over some sort of with resolution in frequency (195.3125 kHz-wide channels), time
“random walk”. Since the strength of the refraction is frequency- (10 s sub-integrations), polarisation (four coherency products)
dependent, this also means that the photons we receive at and rotational phase (1024 phase bins).
different frequencies traverse different parts of the IISM and A few changes to the set-up of this network were made over
therefore may sample regions with different electron densities. In the years during which these observations were taken. Specifi-
principle this would lead to frequency-dependent measurements cally, in August 2013 a new data-taking mode was deployed to
of DM in the case of inhomogeneous media. As demonstrated allow a reduction in the number of bits with which the data were
by Cordes et al. (2016) both theoretically and through simula- recorded (reduced to eight bits from the original 16 bits). Since
tions, the fact that the IISM volumes sampled by the radio waves the total data rate which the recording computers and network
differ across frequencies effectively causes the DM time series links can keep up with limits the total amount of data that can
observed at low frequencies to be very similar to a low-pass- be recorded, this reduction in bits enabled an increase in observ-
filtered (or smoothed) version of the DM time series measured ing bandwidth. Consequently, the observing bandwidth doubled
at higher observing frequencies (see Figs. 3 and 4 of Cordes et al. from 47.7 to 95.3 MHz starting in mid August 2013. This change
2016). in bandwidth also slightly changed the centre frequency of the
In practice, this effect has not been observed, although lim- observations, which moved from 138.77 to 149.90 MHz at that
its have been placed using extremely wide ranges of frequencies time. This implies a shift of the centre frequency by an integer
(Hassall et al. 2012; Pennucci et al. 2014). An observational test number (57) of frequency channels. (The last observation with
of this phenomenon would allow realistic tests of how such DM 47.7 MHz of bandwidth in our data set was taken on 19 August
chromaticity may affect the usefulness of low-frequency DM 2013 while the first with 95.3 MHz of bandwidth was recorded
time series for correcting higher-frequency pulsar-timing data. on 27 August 2013.)
In this paper, we present high-cadence low-frequency obser- For technical reasons, we restricted the bandwidth of observ-
vations of PSR J2219+4754, a slow pulsar discovered by ing to 71.5 MHz from February 2015 onwards. In order to
Taylor & Huguenin (1969) with a DM of 43.5 cm−3 pc. Observed minimise the impact of this bandwidth reduction on the scientific
with the LOw Frequency ARray (LOFAR) telescope, it is one of quality of the data, the observed bandwidth was kept centred on
the brightest sources in the northern sky (see Bilous et al. 2016). the most sensitive part of the bandpass, thereby causing the cen-
Combined with the high fractional bandwidth of LOFAR, this tre frequency to shift slightly from 149.90 to 153.81 MHz. This
leads to a very high DM measurement precision (see Verbiest & shift was again made by an integer number of frequency channels
Shaifullah 2018). Strong DM variations have previously been (20 channels in this case), so that the frequencies of individual
reported for this source by Ahuja et al. (2005). Additionally, channels remained constant over the entire dataset.
this pulsar is known to show profile-shape variations. While
Suleymanova & Shitov (1994) concluded that the profile-shape
variations they observed are intrinsic to the pulsar, more recently, 3. Data analysis
the analysis of our companion paper (Michilli et al. 2018) The data analysis has been carried out using the PSRCHIVE
suggests an interstellar origin. (Hotan et al. 2004; van Straten et al. 2012), TEMPO 2 (Hobbs
In Sect. 2 we describe the observations used in our work, et al. 2006) and COASTGUARD (Lazarus et al. 2016) software
while Sect. 3 explains the steps taken in deriving the DM time packages, as detailed below.
series (and the detected frequency dependence of the measured
DM values). Section 4 discusses the nature of the DM vari-
ations observed and assesses possible implications for pulsar 3.1. Pre-processing
timing. Section 5 concludes the paper by summarising our main Before any of the other analysis steps were carried out, the
findings. data were inspected for man-made radio-frequency interference
(RFI) and any affected channel-subintegration combinations
2. Observations were removed from the data using the “Surgical” algorithm from
the CLEAN . PY script, which is part of the COASTGUARD python
Our analysis is based on data from three German stations of the
International LOFAR Telescope (ILT, van Haarlem et al. 2013), 4 Publicly available at https://github.com/AHorneffer/
namely the stations in Effelsberg (telescope identifier DE601), lump-lofar-und-mpifr-pulsare and described on https://
Tautenburg (DE603) and Jülich (DE605), between 25 February deki.mpifr-bonn.mpg.de/Cooperations/LOFAR/Software/
2013 and 25 November 2016 (see Table 1). LOFAR is described LuMP

A22, page 3 of 11
 A&A 624, A22 (2019)

Table 1. Summary of observations.

Telescope Nobs Gregorian MJD Observation Length Bandwidth Location

identifier date range range range median (MHz)
DE601 17 24/05/2013–08/01/2014 56436–56665 7–60 min 28 min 47.7, 95.3 Effelsberg
DE601 2a 19/05/2015–20/05/2015 57161–57162 9 h, 8 h 95.3 Effelsberg
DE603 10 12/02/2014–03/05/2014 56700–56780 7–13 min 13 min 95.3 Tautenburg
DE605 119 07/03/2014–25/11/2016 56723–57717 2–146 min 115 min 95.3, 71.5 Jülich

Notes. Given are the telescope identifier; the number of observations Nobs ; the time span of the observations; the range of observation lengths
and the median observation length; the bandwidth of the observations (which changed for DE601 and DE605 as discussed in Sect. 2) and the
geographical location of the stations. (a) These observations were omitted from the analysis, and are used as reference template instead.

package5 . In addition to this, throughout the analysis the original to an arbitrary, fully frequency-averaged high-S/N observation,
data from outlier points were also visually inspected to ensure using the PSRCHIVE program PAAS. This analytic template was
the absence of RFI. Where needed, RFI was manually removed cross-correlated against the profiles of each frequency channel of
using the PSRCHIVE program PAZI upon which the processing each observation (using the standard method described by Taylor
was repeated. (The number of data points that were reprocessed 1992), after which the TEMPO 2 software package was used to
in this way was minimal.) Typically about 25% of our data were fit for the DM at each observing epoch. Subsequently these
excised due to presence of RFI. This percentage varied depend- daily DM measurements were held fixed in the timing model of
ing on the time of day of the observation (with higher prevalence Hobbs et al. (2004) while the entire data set was used to fit the
of RFI during the day and lower during the night) but no sig- pulsar’s spin period, spin period derivative and position. This
nificant difference was seen in the RFI fraction of the different updated timing model was then used to re-do the time-averaging
stations. of the observations, after which the process was iterated until the
The data were calibrated in polarisation following the meth- timing model and DM time series converged.
ods outlined in Noutsos et al. (2015), after which the coherency After this initial timing analysis, the proper motion in the
products were combined to yield total intensity. Given the lim- timing model was updated from the values of Lyne et al. (1982)
ited length of the observations (typically shorter than two hours) to those of Michilli et al. (2018; which was not available at the
and the fact that the hour angle and elevation of all observations start of our analysis), although this had no significant impact
were highly similar as the observations were always scheduled on our results given the short time-span of our observations.
close to or across transit, the calibration (and therefore also its Since this pulsar exhibits large amounts of timing noise (Hobbs
imperfections) did not significantly affect our analysis. Next, et al. 2004) and because our timing analysis fully ignored
the total-intensity profiles were averaged in time, resulting in frequency-dependent profile evolution and the temporal evolu-
a single, frequency and phase-resolved pulse profile for every tion of the profile described in our companion paper (Michilli
observation. et al. 2018), the results of our timing analysis were not ideal and
The data were not averaged in frequency, but in order to pro- certainly not predictive enough to warrant publication of the tim-
duce as homogeneous a data set as possible, the data observed ing model. However, the timing model thus obtained did succeed
with 95.3 MHz of bandwidth were downsized to 71.5 MHz in phase-aligning our observations to within ∼400 µs, that is,
instead, by cutting out the edges of the band, where the sensi- within a phase bin; and much more precisely within any given
tivity is low due to the presence of filters (van Haarlem et al. observation. We therefore used this timing model for the final
2013). The full-bandwidth data remains available upon request time-averaging of our data.
for possible follow-up investigations, when needed. In order to determine highly precise and reliable DM val-
ues, a more advanced timing analysis was carried out. For this
analysis we used a frequency-resolved template. To create this
3.2. Timing and DM time series
template we combined two long observations taken with DE601
In order to determine the DM time series, we used the pulsar- on 19 May 2015 (MJD 57161), from 03:00 to 12:00 UTC and
timing technique (see Lorimer & Kramer 2005). While an on 20 May 2015 (MJD 57162), from 00:00 to 07:50 UTC,
accurate timing model is in principle not required for instanta- for a total effective duration of 16.6 h. This observation was
neous measurements of DM, the time-averaging of observations averaged in time and summed to total intensity, providing a
does improve if the pulsar timing model is of good quality. frequency-resolved pulse profile with a S/N a few times that of
Consequently we carried out an initial, straightforward timing the typical observation. This template was subsequently used as
analysis based on a single analytic template profile constructed the phase reference for timing and was otherwise fully omitted
of von Mises functions (see, e.g. Jammalamadaka & SenGupta from our analysis. The pulse times-of-arrival (ToAs) were in this
2001) of the form: case determined using the Fourier-Domain Monte-Carlo (FDM)
approach6 , as advised by Verbiest et al. (2016), on a channel-
f (x) = A · eκ(cos 2π(x−µ)−1) (8) by-channel basis (i.e. resulting in up to 366 simultaneous ToAs
per observation). Since the template profile is data-derived
with A being the amplitude of the component, κ the so-called and frequency-resolved, any static frequency dependence of the
compactness, and µ the pulse phase. One full rotation corre-
sponds to one unit in x. Around five of these functions were fitted
6 This algorithm is identical to that described by Taylor (1992), except
5 Publicly available at https://github.com/plazar/coast_ for the uncertainties. FDM uses either formal uncertainties or a Monte-
guard Carlo simulation. We used the default formal uncertainties.

A22, page 4 of 11
 J. Y. Donner et al.: First detection of frequency-dependent, time-variable dispersion measures

Year
MJD 56570
2013.3 2013.8 2014.4 2014.9 2015.5 2016.0 2016.6 2017.1 1 MJD 57161
4 8.5
0.8

Normalised flux
3 6.4
∆DM (10-3cm-3pc)

∆t at 1.4GHz (µs)
2 4.2 0.6
MJD 57161
MJD 56570
1 2.1
0.4
0 0.0
0.2
-1 -2.1

-2 -4.2 0
0.36 0.38 0.4 0.42 0.44 0.46
56400 56600 56800 57000 57200 57400 57600 57800 Pulse phase
MJD
Fig. 2. Peak-aligned pulse profiles for a scattered observation on
Fig. 1. DM variations in the direction of PSR J2219+4754. The vertical MJD 56570 and an unscattered observation on MJD 57161, which is part
line indicates the time at which the last change to the centre frequency of the standard template. The profiles were aligned to have the leading
and bandwidth of our data occurred, i.e. all data to the right of the edge and peak (which have identical shape at both dates) to align; and
dashed line have identical bandwidth and centre frequency (except for the pulses shown are integrated over the full bandwidth of 71.5 MHz.
slight deviations due to variations in RFI excision). The two arrows
indicate the observations that were used to quantify the effect of scat-
tering (see Sect. 3.3), with the latter of these two (MJD 57161) also being more pronounced at lower frequencies than at higher fre-
being the standard template. A DM baseline of 43.48205 cm−3 pc has
been subtracted. The additional error bars on the lower end of the
quencies. The net effect of such additional profile components on
pre-MJD 57000 data points indicate the expected impact of profile our DM measurements would be to delay the measured ToAs;
scattering, as described and quantified in Sect. 3.3. The second y-axis and since the additional components are more pronounced at
indicates the corresponding dispersive delay at an observing frequency lower frequencies, this would lead to an overestimate of the DM.
of 1.4 GHz. Figure 2 shows the worst-case profile-shape variability7 present
in our data set, comparing the pulse profile at MJD 56570, when
the scattering components are most pronounced, to the standard
template shape would not affect our analysis, as it is inher-
template from MJD 57161 when the scattering is minimal, but
ently taken into account. (Also, any DM measurement derived
some reflected “echo” images of the main pulse are faintly vis-
from this analysis is by definition referred to the DM incor-
ible at longer lags (see Michilli et al. 2018, for a full discussion
porated in this template, thereby rendering an absolute DM
of these pulse-shape variations).
measurement impossible.) However, time variability of the pulse
To quantify the impact of the scattered power in the trailing
profile shape (as investigated by Michilli et al. 2018), does have
edge of the pulse profile shown in Fig. 2, an analytic model of
the potential to negatively impact the reliability of our results.
the (unscattered) template profile was extended with extra power
This is discussed in detail in Sect. 3.3.
in the trailing edge to match the one shown in the observation
In this more advanced timing analysis we do not fit any time-
of MJD 56570. By timing both of these analytic pulse models
dependent timing-model parameter but exclusively for DM on an
(the one with additional power in the trailing edge and the one
observation-by-observation basis. Since these DM fits are based
without) and repeating the process at three different frequencies,
on a set of simultaneous ToAs (i.e. those derived from the dif-
we could directly determine the impact the scattering has on the
ferent channels of the given observation), no time-dependent
ToAs at different frequencies.
timing-model parameters affect our results. An arbitrary phase
Fitting the DM for the scattered and unscattered profiles
offset is routinely subtracted along with every fit, in order
returns a difference of ∆DM = 1.5(4) × 10−4 cm−3 pc. We note
to prevent biases discussed by Keith et al. (2013). The DM
this DM difference is the maximum offset induced by profile-
measurements from this analysis had a median uncertainty of
shape changes. A similar test for the reflected “echo” images
3.7 × 10−5 cm−3 pc and their time series is presented in Fig. 1,
visible at later MJDs lead to an insignificant impact on the DM.
but before deeper consideration of these DM values, some
In comparison to the DM variations and DM measurement pre-
corrupting influences will be discussed below.
cision shown in Fig. 1, we note that enhanced scattering by this
amount does slightly alter the results, but does not fundamen-
3.3. Impact of pulse-shape variations on the DMs tally change the shape of the (much stronger) DM variations we
identified.
As mentioned in the previous paragraph, time variations in Now that the maximum impact of the profile-shape vari-
the shape of the pulse profile could corrupt our measured DM ations on our DM measurements has been established, we
values, particularly if this time variability is also frequency investigate how the amplitude of these profile-shape differences
dependent. The profile-shape evolution discussed in our com-
panion paper (Michilli et al. 2018) is of particular concern 7 As quantified by the goodness-of-fit of the template cross-correlation
since it is known to affect our data and has been shown to be during the timing with the FDM algorithm described in the previous
frequency dependent. Specifically, Michilli et al. (2018) report section. While the scattering strength does appear to be marginally
time-variable scattering that shows up as additional pulsed com- stronger at earlier dates, the narrower bandwidth of the earlier obser-
ponents at and near the trailing edge of the pulse profile. As vations cause the profile-shape difference to be partly covered up by
scattering is frequency dependent, so are these components, increased levels of radiometer noise.

A22, page 5 of 11
 A&A 624, A22 (2019)

Time (sec) Year

2013.3 2013.8 2014.4 2014.9 2015.5 2016.0 2016.6 2017.1
-0.01 0 0.01 0.02

Extrema of residuals (peak normalised)

0.1 Maximum residual
Minimum residual
56400
0.05

56600
0

56800 -0.05
56400 56600 56800 57000 57200 57400 57600 57800
MJD, profile residuals

MJD

57000 Fig. 4. Maxima and minima of the profile-shape differences shown

in Fig. 3. Error bars indicate 2.88 times the off-pulse RMS (since for
Gaussian noise there is a 1/256 chance of a measurement falling more
than 2.88σ away from the mean of the distribution and the pulse pro-
file differences considered here consist of 256 bins). The profile-shape
57200 distortion due to scattering is clearly visible before MJD 57000 with a
residual amplitude that decays linearly in time. Between MJDs 57000
and 57600 the profile differences are not fully consistent with Gaus-
sian noise, but are stable. After MJD 57600 some further, lower-level,
variations occur.
57400

profile, the scattering reaches a minimum and the pulse-profile

is stable henceforth, with the exception of radiometer noise and
57600 the lower-lying variations documented by Michilli et al. (2018).
To more quantitatively evaluate this evolution, the maximum and
minimum values of these profile-shape differences are plotted as
a function of time in Fig. 4. Here the linear decay of the pro-
file residuals between the start of our data set (MJD 56436) and
57800
0.14 0.16 0.18 0.2 MJD 57000 is clearly seen. Noting that the DM impact of the
worst-scattered profile (as quantified above) is 1.5 × 10−4 cm−3 pc
Pulse phase and that the amplitudes of any profile-shape variations after
MJD 57000 are lower by at least a factor of five (or more), we
Fig. 3. Pulse-shape variations for PSR J2219+4754. Shown are the do not expect DM corruptions at levels beyond 10−4 cm−3 pc
profile-shape differences between the observations and the standard beyond this date. For dates before MJD 57000 we consider a
template, which is indicated by the dashed horizontal line at MJD 57161. potential DM overestimation with amplitude 1.5 × 10−4 cm−3 pc
The vertical line indicates the position of the peak of the pulse profile.
For a clearer view, only one observation is plotted for every four-week
at MJD 56570 and linearly decreasing to zero by MJD 57000.
interval. The pulse profiles used for this plot were fully averaged in fre-
quency and downsampled to 256 profile bins (to reduce the radiometer 3.4. Frequency dependence of DM
noise in the plot). The differences between observations and standard
template were calculated after peak-normalising and aligning the pro- The DM time series described earlier (Fig. 1, Sect. 3.2) was
files by their leading edge and peak. This figure is consistent with Fig. 4 derived by carrying out a standard least-squares fit to ToAs
of Michilli et al. (2018), who used a slightly extended dataset. from the various frequency channels of any given observation,
using the TEMPO 2 pulsar-timing software. During these fits, the
median reduced χ2 value was 2.62, indicating that either the
evolves in time. Specifically we are concerned with the ampli- input ToA uncertainties were underestimated, or that unmodelled
tude variations of the scattered power in the trailing edge that (i.e. frequency-dependent) structure was present in the post-fit
is visible primarily at the earlier epochs: MJD 56400–57000 at timing residuals. Non-unity reduced χ2 values are not abnormal
pulse phase elongations of ≤2% from the pulse peak. This is in in pulsar timing since several reasons for inaccurate estimation
contrast to the lower-lying “echo” images analysed by Michilli of ToA errors could be present (see Verbiest & Shaifullah 2018,
et al. (2018), which have much lower amplitude (and conse- for an extensive review). Typically, however, these effects only
quently less impact on timing or DM measurements) and lie at cause reduced χ2 values that are lower than two (Verbiest et al.
pulse longitudes further away from the pulse peak (>2% of a 2016). A more astrophysical potential cause, which is expected
pulse period). to be most pronounced at low frequencies, is chromaticity of the
Figure 3 shows the profile-shape differences of our obser- observed DMs.
vations with respect to the template observation. This clearly As described in Sect. 3.2, since our timing is based on a
shows the excess power in the trailing edge at early epochs, with data-derived template, we are only sensitive to differences in DM
an amplitude that decays in time. At the epoch of the template between the observation and the template observation and hence
A22, page 6 of 11
 J. Y. Donner et al.: First detection of frequency-dependent, time-variable dispersion measures

Year
2013.3 2013.8 2014.4 2014.9 2015.5 2016.0 2016.6 2017.1
4 > 149 MHz 8.5
< 149 MHz
3 6.4
∆DM (10-3cm-3pc)

2 4.2
1 2.1

∆t at 1.4GHz (µs)
0 0.0
-1 -2.1
-2 -4.2

2 4.2
DMlow - DMhigh

1 2.1
0 0.0
-1 -2.1
56400 56600 56800 57000 57200 57400 57600 57800
MJD
Fig. 5. Top panel: DM time series for the upper and lower halves of the observing band. We note that until MJD 56524, the observing bandwidth
was only 47 MHz, which leads to worse DM precision and a smaller DM difference. Bottom panel: difference between the DM measured in the
top and bottom part of the band. By definition (i.e. through selection of the standard profile) no DM difference is present at MJD 57161. While
the long-term DM gradient appears consistent between the two bands, the higher-frequency DM structures are shown to affect the high-frequency
band more than the low-frequency band. The second y-axis in both panels indicates the corresponding dispersive delay at an observing frequency
of 1.4 GHz.

mere “DM chromaticity” would not be visible in our analysis. at a given time lag τ is derived from the DM time series using
However, if this frequency-dependent DM would be variable the following equation:
in time (as might be expected given the significant changes in
the overall DM), then any non-ν−2 dispersive effects should be DDM (τ) = h[DM(t + τ) − DM(t)]2 i (9)
equally time-dependent.
Figure 5 shows the time series for the DMs measured from using a weighted mean (i.e. the ∆DM values were weighted by
the top part (149–190 MHz, centred at 169 MHz) and bottom part 1/(σ2DM(t+τ) + σ2DM(t) )). The uncertainties of DDM (τ) are derived
(118–149 MHz, centred at 133 MHz) of our band, along with the through Monte-Carlo simulations, by varying the DM time series
difference between these. The frequency dependence of these according to the DM measurement uncertainties and thereby
DM measurements is highly significant and exceeds the maximal identifying the 68% confidence intervals of DDM (τ) over 10 000
impact of the profile-shape variability, as quantified in Sect. 3.3, simulations.
by an order of magnitude. Figure 6 shows the structure function of the DM time series
presented in Figs. 1 and 5. Due to the high observing cadence
3.5. Structure function analysis with the GLOW stations, the shortest lags we sample in our
structure function go down to a few days. Previously published
Given the strong and significant DM variations we detected – as structure functions of DM time series sampled time lags down to
well as the frequency dependence of these measurements – it is tens of days (You et al. 2007; Jones et al. 2017) or about 100 days
worthwhile to evaluate the overall structure of the IISM towards (Keith et al. 2013), so with our dataset we extend the range at
this pulsar in order to verify whether the DM variability could which Kolmogorov turbulence has been tested in DM time series
be explained by standard IISM turbulence or not. To this end, by one order of magnitude.
we compute the structure functions of the DM variations shown We find that the structure function of the DM time series
in Figs. 1 and 5 and compare these to a Kolmogorov turbulence agrees extremely well with a Kolmogorov spectrum. This con-
density spectrum, which is known to usually be a good approx- formity seems to corroborate the findings by Pen & King (2012);
imation for the IISM density spectrum in general (Armstrong Pen & Levin (2014) that ESEs may actually be part of the larger-
et al. 1995; Keith et al. 2013), although deviations have been scale IISM turbulent structure rather than separate, localised,
reported (see Gupta 2000, for a review). The structure function events.
A22, page 7 of 11
 A&A 624, A22 (2019)

Estimated spatial scale (AU) companion paper (Michilli et al. 2018) are consistent with a
0.07 0.7 7 70 number of contained, refractive lenses near the line of sight. Nev-
1e-04
Full band
ertheless, the structure function (Fig. 6) is fully consistent with
< 149 MHz a Kolmogorov spectrum. In the following, we will consider a
1e-05
> 149 MHz
Kolmogorov
simplistic model based on three individual, spherical interstellar
lenses in order to compare the required lens sizes and densities to
those previously published for ESEs9 . We presume these lenses
1e-06
to be the cause of the DM excess around MJDs 56600 and 57000.
DDM(τ)

The steep drop-off in DM towards the end of our data set will be
1e-07 disregarded as further monitoring of this pulsar’s DM time series
is required for this event. (Particularly the apparently third ESE
which seems to only affect the lower-frequency part of our data
1e-08
deserves further analysis and continued observations.)
In order to identify the presumed ESEs clearly, we define a
1e-09 baseline DM level of 43.482 cm−3 pc, which is the average DM
1 10 100 1000 value from MJD 57100 to MJD 57200. This value is chosen as
Time interval τ (days)
reference because it describes the only time window in our data
where no variations in DM are observed. It is also used as the
Fig. 6. Structure functions of the DM time series from Figs. 1 and 5. The
spatial scale is estimated for turbulence half-way to the pulsar, using the
reference value for all plots. The reference observation against
proper motion published by Michilli et al. (2018). The Kolmogorov tur- which the timing was performed was also selected from within
bulence model is characterised by the structure functions of simulated this MJD range. We furthermore only consider the DM peak
DM time series derived from a power spectrum with spectral index 5/3. near MJD 57000 because this is the most easily identified and
The dotted lines represent the sample average and 1σ-contours of 1000 the differences with the other potential ESEs are well within the
iterations. The amplitude of the Kolmogorov model was scaled to fit the uncertainties of our model, so similar results can be considered
first 200 days of the data, because the structure function becomes very to hold for all three presumed clouds.
uncertain at the longest lags. We will now model the aforementioned DM peak, assuming
a spherical, homogeneous cloud of ionised gas to be causing it.
Using Eq. (28) of Lam et al. (2016) for Kolmogorov tur- The time from the maximum to the end of the peak is 150 days,
bulence, neglecting Earth’s motion and assuming a constant so the total duration of the cloud passage is estimated to be
amplitude of the electron density power-law spectrum (C N2 ) 300 days. Using a proper motion of 22.2 mas yr−1 (see Michilli
along the ling of sight, the latter can be directly related to the et al. 2018), the angular size of the cloud can be calculated as
amplitude of the DM structure function: θ = 18 mas. To estimate the physical size, the distance to the
cloud is needed. This distance can only be estimated, but it has
3 8/3 5/3 5/3 to be lower than the distance to the pulsar, which is estimated to
DDM (τ) = C N2 × 88.3 × z µ τ , (10)
8 p be 2.2 kpc (Cordes & Lazio 2002). Combining the angular size
with the proper motion µ and the pulsar distance zp . From the of the cloud and the distance to the pulsar, the maximum size of
Kolmogorov power-law fit to the structure function for time the cloud (if it was directly in front of the pulsar) is 40 AU.
scales up to 200 days, we get Given the assumption of a spherical object, the maximum
path length through the cloud is equal to its lateral extent,
 τ 5/3 and from this path length and the maximum DM increase one
DDM (τ) = 3.1 × 10−10 cm−6 pc2 . (11) can calculate the average excess electron density in the cloud.
1d
From the upper limit on the cloud size follows a lower limit on
With a proper motion of 22.2 mas yr−1 (see Michilli et al. 2018) the average extra electron density, which is 15 cm−3 . Compared
and a pulsar distance of 2200 pc (Cordes & Lazio 2002)8 , we to a typical electron density in the Warm Ionised Medium of
get C N2 = 0.9 × 10−3 m−6.67 . This is in close agreement with the 0.1–0.5 cm−3 (Haverkorn & Spangler 2013), our model suggests
findings of Armstrong et al. (1995), that C N2 = 10−3 m−6.67 fits a an overdensity of about two orders of magnitude. The solid lines
huge range of spatial scales and thus supports the idea that the in Fig. 7 show the estimated cloud size and its electron density
observed variability is part of the general IISM turbulence rather depending on its distance to the Earth. Some values from the lit-
than a stand-alone ESE. erature of similar estimates are added for comparison: the first
ESE observed by Fiedler et al. (1987, without distance estimate),
an ESE observed in flux density and timing residuals modelled as
4. Discussion two clouds of identical size by Cognard et al. (1993), the three-
4.1. Origin of the DM variability year-long ESE observed in flux density by Maitia et al. (2003,
without uncertainties), and the models of two ESEs observed by
Both the amplitude and shape of the DM variations shown Coles et al. (2015). These latter two ESEs were rescaled to be
in Fig. 1 are reminiscent of ESEs presented elsewhere in the compatible with pulsar distances based on the parallax measure-
literature (e.g. Cognard et al. 1993; Coles et al. 2015) and inter- ments of Ng et al. (2014) and Reardon et al. (2016), corrected
preted as individual lenses of ionised matter. In addition to this, for the Lutz-Kelker bias following the analysis by Verbiest et al.
the scattering events seen in this pulsar and discussed in our (2012) as corrected by Igoshev et al. (2016).
8 https://www.nrl.navy.mil/rsd/RORF/ne2001/. Recently,
Yao et al. (2017) published a new Galactic electron density model 9 We cannot analyse potential correlations between scintillation and
(http://119.78.162.254/dmodel/), which gives a consistent DM variability like Coles et al. (2015) did because the scintillation
distance estimate of 2.4 kpc. bandwidth is smaller than our frequency resolution.

A22, page 8 of 11
 J. Y. Donner et al.: First detection of frequency-dependent, time-variable dispersion measures

1000 we similarly insert Eq. (4) into Eq. (7) and solve for Bk = B cos γ:
Cognard et al. (1993)
Cloud density (cm-3)

Maitia et al. (2003)

100
ne 1 4 ν12 ν22
Cognard et al. (1993)
= × T 1 = 1.81 × 108 T 1 , (12)
10
Coles et al. (2015) cm−3 (8.98 kHz)2 3 ν12 + ν22
Fiedler et al. (1987)
Bk 1 1 ν2 − ν12
1 = × T 2 ν1 ν2 23 = 1.76 × 101 T 2 . (13)
60 1G 2.80 MHz 2 ν2 − ν13
Maitia et al. (2003)
Coles et al. (2015)
Cloud size (AU)

40 T 1 or T 2 on the order of 5e-5 would imply variations of the

electron density of the order of 104 cm−3 , or variations of the
20
Fiedler et al. (1987)
magnetic field along the line of sight on the order of 10−3 G.
0 Coles et al. (2015)
This is equivalent to the DM and RM varying by many orders of
Cognard et al. (1993)
magnitude of their actual value. As shown in this paper, the DM
0 0.5 1 1.5
Distance (kpc)
2 2.5 3
only changes by fractions of its total value. A detailed analysis of
the RMs in these data is in progress and will be published in due
Fig. 7. Estimated cloud size and electron density for the second compo- course; but any RM variations towards this source are minimal
nent of the ESE depending on its distance to the Earth, represented as and would not satisfy this scenario, either.
solid lines in the plot. Horizontal lines represent literature values with- As DM chromaticity is usually not observed, a localised
out distance estimates. The vertical line represents the distance estimate structure as the origin seems far more likely. If a structure cor-
for the structures that cause the scattering “echoes” from our companion responding to a DM excess on the order of 10−2 cm−3 pc was
paper (Michilli et al. 2018). The cloud size estimates of Cognard et al.
causing the observed variations of the frequency-dependence (as
(1993) for the two clouds forming the ESE are very similar (0.050 and
0.094 AU) and thus indistinguishable in the plot. The error bars on the expected from the DM time series in Fig. 1), this would require
cloud sizes of Coles et al. (2015) represent the non-spherical character variations in T 1 + T 2 of the order of 0.2, implying electron den-
of the clouds and not the actual uncertainty of the cloud size. sity variations of the order of 4 × 107 cm−3 , or variations of the
magnetic field strength of the order of 3 G. To only contribute
10−2 cm−3 pc to the DM, the structures would need to be 5e-5 AU
When comparing the size and density estimates to typical thick, which is in strong contradiction with the long duration of
values from the literature, this particular ESE could be anywhere these variations. Specifically, a 5e-5 AU-thick region with an
along our line of sight without falling out of the sample, as the electron-density excess of the order 107 cm−3 could be gener-
densities and sizes in the literature span the entire range of possi- ated by a chance alignment of a star with the line of sight, but
ble values (with an exception for clouds that are very close to the given the rapid spatial motion of the line of sight, such an align-
Earth as the implied density would become unrealistically high). ment would pass quickly. Michilli et al. (2018) also discussed the
In particular a cloud roughly halfway to the pulsar (in agreement potential impact of stellar winds on the observations of this pul-
with Michilli et al. 2018) with a size of about 20 AU and a density sar, but in that scenario the offset of the relevant star to the line of
of a few tens of electrons per cm3 would be highly comparable to sight is too far to allow the mechanism proposed by Tanenbaum
previously observed ESEs. All calculations for this simple model et al. (1968) as an explanation for the frequency dependence of
are only very rough estimates, as it is impossible to disentan- the DM since the size and density of the stellar-wind bubble
gle the different components of the DM time series and to find along the line of sight would not match the predictions derived
the correct DM baseline. It is, for example, highly likely that above.
there are multiple clouds of various sizes, which overlap (see As these scenarios are highly unrealistic, we conclude that
also the discussion by Michilli et al. 2018). The steepest DM deviations from the dispersion law as described by Tanenbaum
decrease, which starts around MJD 57000, can be interpreted as et al. (1968) are not the cause of the frequency-dependent DMs
the edge of one cloud and lasts about 75 days. Estimating the we detected. We therefore favour the explanation from Cordes
cloud to be located half-way to the pulsar, this edge is about et al. (2016), that refractive effects lead to a frequency depen-
5 AU thick. This is roughly of the order of what Brisken et al. dence of the medium the pulsar radiation passes through. While
(2010) found for elongated filaments, so it could be the case that the DM time series at 133 MHz does on the whole look smoother
the filaments they see are actually the edges of ionised clouds than the variations at 169 MHz (as expected from the analysis by
(see also Pen & Levin 2014; Liu et al. 2016). Cordes et al. 2016), at a few epochs (e.g. around MJD 57100 and
shortly after MJD 57200) the lower frequencies show more dra-
4.2. Origin of the DM frequency dependence matic DM trends, which may be in tension with the theoretical
expectations of this model.
In the following, we will assume that the observed DM differ-
ence was caused by deviations from the dispersion law (Eq. (1)) 4.3. Consequences for high-precision pulsar timing
due to invalidity of the assumptions that ν  νp or ν  νc .
The observed variability of ∆DM on the order of Variations in the DM that cannot be accurately and precisely
10−3 cm−3 pc between our two bands centred at 133 and 169 MHz measured and modelled are a problem for pulsar timing, as they
corresponds to an additional time delay of the order of ∼200 µs, add a time-dependent extra delay to the ToAs. If the DM vari-
which is smaller than the total dispersive delay by a factor of ations we reported here were to occur along the line of sight to
about 5e-5, so the variability of the sum of T 1 and T 2 would pulsars used in high-precision timing experiments, they would
have to be of the order of 5e-5 as well (see Eq. (5)) if we assume corrupt astrophysically relevant parameters and would signifi-
a homogeneous medium along the line of sight. To relate T 1 to cantly reduce sensitivity to interesting signals (see, e.g. You et al.
an electron density, we insert Eq. (3) into Eq. (6) and solve for ne . 2007) if their impact on the data was not removed by, for exam-
To relate T 2 to a magnetic field strength along the line of sight, ple, measuring the DM at every epoch and correcting for it. This
A22, page 9 of 11
 A&A 624, A22 (2019)

correction can only be done with simultaneous multi-frequency or proper motion of the IISM structures and the DM baseline
or low-frequency data, which are not always available (see e.g. level. The simple spherical model discussed in this paper does
the first IPTA release, Verbiest et al. 2016). In the potential ESE not provide definite values, but does provide limits on the poten-
discussed in this paper, the maximum difference in the extra tial ESE’s electron density and size, which are of the same order
time-delay across the entire dataset (caused by a DM difference of magnitude as previous results for ESEs. Further observations
of 6 × 10−3 cm−3 pc) at the commonly used 21 cm wavelength of ESEs will help to improve the constraints on the size, because
would imply structures of the order of 13 µs in the timing resid- the lack of knowledge of the distance to the IISM structures is
uals. This is well above the precision needed for high-precision less of an issue when a larger sample of observations is given,
pulsar timing experiments, which require sub-microsecond pre- assuming a homogeneous distribution of ESEs in the Galaxy. We
cision (see, e.g. Jenet et al. 2005). The ToA difference across a furthermore point out that the variations presented here could
250 MHz bandwidth centred at 1.4 GHz due to the extra disper- well be part of a uniform spectrum rather than separate, distinct,
sion is 2 µs, so the ToA precision has to be substantially better structures.
to properly measure and correct the impact the ESE has on the Finally, we have presented the first observational evidence
data (we note that the median timing precision in the IPTA is for frequency-dependent DMs and have confirmed that the long-
currently 2.5 µs, Verbiest et al. 2016). As shown by Lee et al. term DM trends are consistent across the frequencies we probed,
(2014, Eq. (12)), the precision of this correction would however whereas the shorter-term DM structure is highly chromatic. This
be about an order of magnitude worse than the ToA precision and bodes well for efforts to apply low-frequency DM time series
averaging the DM values to increase precision is typically not as corrections to high-frequency pulsar-timing data, although
a valid solution because of the usually low sampling rate (see further study across a wider range of frequencies should be
Verbiest et al. 2016, and references therein) and possible short undertaken to quantify any potential corruptions such correc-
time scales of the variations. Thus, correcting high-frequency tions would cause.
observations with DM values measured from that same observa-
tion would not suffice to correct for DM variations similar to the Acknowledgements. J.Y.D. acknowledges financial support during part of
ones presented in this paper. this work from a Brückenstipendium granted by Bielefeld University. S.O.
acknowledges the support from the Alexander von Humboldt Foundation and
While low-frequency data are very useful in computing the Australian Research Council grant Laureate Fellowship FL150100148.
highly detailed measurements of DM variations, the chromatic- J.W.T.H. and D.M. acknowledge support from the European Research Council
ity we have presented does cause concern as it may imply a under the European Union’s Seventh Framework Programme (FP/2007-2013) /
mismatch between the DM values observed at lower frequencies ERC Grant Agreement nr. 337062 (“DRAGNET”). This paper is based on
data from the German LOng-Wavelength (GLOW) array, which is part of the
and those observed at higher frequencies. We note, however, that International LOFAR Telescope (ILT) which is designed and built by ASTRON
chromaticity mostly perturbs DM variations on short timescales, (van Haarlem et al. 2013). Specifically, we used the Effelsberg (DE601) station
whereas the long-term DM trends tend to show reasonable levels funded by the Max-Planck-Gesellschaft; the Tautenburg (DE603) station funded
of agreement (as suggested by Cordes et al. 2016, and con- by the State of Thuringia and supported by the European Union (EFRE) and
firmed by our analysis, see the bottom panel of Fig. 5). Since the Federal Ministry of Education and Research (BMBF) Verbundforschung
project D-LOFAR I (grant number 05A08ST1); and the Jülich (DE605) station
the spatial electron-density spectrum in our Galaxy (and there- supported by the BMBF Verbundforschung project D-LOFAR I (grant number
fore the spectrum of DM variations) has more power at lower 05A08LJ1). The observations of the German LOFAR stations were carried out
frequencies (Armstrong et al. 1995), it is particularly the longer- in the stand-alone GLOW mode which is technically operated and supported
term DM variations that require correction in high-frequency by the Max-Planck-Institut für Radioastronomie, the Forschungszentrum Jülich,
Bielefeld University, by BMBF Verbundforschung project D-LOFAR III (grant
pulsar-timing data, which implies the impact of chromaticity number 05A14PBA) and by the states of Nordrhein-Westfalen and Hamburg.
may be limited. Further studies of chromatic DM variations in The observations during this project were made during station-owners time
pulsars like PSR J2219+4754 – particularly studies that extend as well as during ILT time allocated under project codes LC0_014, LC1_048,
our present analysis to include a wider range of observing LC2_011, LC3_029, LC4_025 and LT5_001. This project also benefited greatly
frequencies – should allow more conclusive answers to these from the ILT Core observations that were included in the Michilli et al. (2018)
paper and which guided some of the analysis and interpretation in this paper.
questions. Those observations were carried out under ILT time allocation codes LC0_011,
LC1_027, LC2_010, LT3_001, LC4_004 and LT5_003. Plots and basic model
fits were created with GNUPLOT.
5. Conclusions
We have presented strong and rapid DM variations along the line
of sight towards PSR J2219+4754, which have a similar ampli- References
tude as the variations commonly seen in millisecond pulsars (see, Ahuja, A. L., Gupta, Y., Mitra, D., & Kembhavi, A. K. 2005, MNRAS, 357, 1013
e.g. Keith et al. 2013). The variations we reported may be caused Armstrong, J. W., Rickett, B. J., & Spangler, S. R. 1995, ApJ, 443, 209
by a group of interstellar clouds typically referred to as ESEs. Bignall, H. E., Croft, S., Hovatta, T., et al. 2015, Advancing Astrophysics with
The ESEs in this paper would be some of the longest and most the Square Kilometre Array (Macclesfield: SKA Ogranisation), 58
Bilous, A. V., Kondratiev, V. I., Kramer, M., et al. 2016, A&A, 591, A134
persistent observed to date. This would also be the first time an Brisken, W. F., Macquart, J.-P., Gao, J. J., et al. 2010, ApJ, 708, 232
ESE is observed in electron density and scattering at the same Cognard, I., Bourgois, G., Lestrade, J.-F., et al. 1993, Nature, 366, 320
time, although the scattering may well be caused by different Coles, W. A., Kerr, M., Shannon, R. M., et al. 2015, ApJ, 808, 113
IISM structures that have only a minimal impact on the DM Cordes, J. M., & Lazio, T. J. W. 2002 ArXiv e-prints
(for the quantitative analysis of the scattering, see the companion [arXiv:astro-ph/0207156]
Cordes, J. M., Shannon, R. M., & Stinebring, D. R. 2016, ApJ, 817, 16
paper by Michilli et al. 2018). Fiedler, R. L., Dennison, B., Johnston, K. J., & Hewish, A. 1987, Nature, 326,
Our frequent observations with the international LOFAR sta- 675
tions allow detailed and highly precise monitoring of the DM Fiedler, R., Dennison, K., Johnston, J., Waltman, E., & Simon, R. 1994, ApJ,
time evolution. The high measurement precision makes us sen- 430, 581
Gunn, J. E., & Ostriker, J. P. 1970, ApJ, 160, 979
sitive to details, which complicates any efforts to provide an Gupta, Y. 2000, in Sources and Scintillations: Refraction and Scattering in Radio
accurate model for the underlying IISM structures. Additionally, Astronomy, IAU Colloquium 182, eds. R. Strom, B. Peng, M. Walker, &
there is a large number of unknown parameters like the distance R. Nan (Dordrecht: Kluwer Academic Publishers), 25

A22, page 10 of 11
 J. Y. Donner et al.: First detection of frequency-dependent, time-variable dispersion measures

Hassall, T. E., Stappers, B. W., Hessels, J. W. T., et al. 2012, A&A, 543, A66 Lyne, A. G., & Lorimer, D. R. 1994, Nature, 369, 127
Haverkorn, M., & Spangler, S. R. 2013, Space Sci. Rev., 178, 483 Lyne, A. G., Anderson, B., & Salter, M. J. 1982, MNRAS, 201, 503
Hewish, A., Bell, S. J., Pilkington, J. D. H., Scott, P. F., & Collins, R. A. 1968, Maitia, V., Lestrade, J.-F., & Cognard, I. 2003, ApJ, 582, 972
Nature, 217, 709 Michilli, D., Hessels, J. W. T., Donner, J. Y., et al. 2018, MNRAS, 476, 2704
Hobbs, G., Lyne, A. G., Kramer, M., Martin, C. E., & Jordan, C. 2004, MNRAS, Ng, C., Bailes, M., Bates, S. D., et al. 2014, MNRAS, 439, 1865
353, 1311 Noutsos, A., Sobey, C., Kondratiev, V. I., et al. 2015, A&A, 576, A62
Hobbs, G. B., Edwards, R. T., & Manchester, R. N. 2006, MNRAS, 369, 655 Pen, U.-L., & King, L. 2012, MNRAS, 421, L132
Hobbs, G., Coles, W., Manchester, R. N., et al. 2012, MNRAS, 427, 2780 Pen, U.-L., & Levin, Y. 2014, MNRAS, 442, 3338
Hotan, A. W., van Straten, W., & Manchester, R. N. 2004, PASA, 21, 302 Pennucci, T. T., Demorest, P. B., & Ransom, S. M. 2014, ApJ, 790, 93
Igoshev, A., Verbunt, F., & Cator, E. 2016, A&A, 591, A123 Rawley, L. A., Taylor, J. H., & Davis, M. M. 1988, ApJ, 326, 947
Jammalamadaka, S. R., & SenGupta, A. 2001, Topics in Circular Statistics (New Reardon, D. J., Hobbs, G., Coles, W., et al. 2016, MNRAS, 455, 1751
Jersey: World Scientific) Romani, R. W., Blandford, R. D., & Cordes, J. M. 1987, Nature, 328, 324
Janssen, G., Hobbs, G., McLaughlin, M., et al. 2015, Advancing Astrophysics Stappers, B. W., Hessels, J. W. T., Alexov, A., et al. 2011, A&A, 530, A80
with the Square Kilometre Array (Macclesfield: SKA Ogranisation), 37 Stinebring, D. R., McLaughlin, M. A., Cordes, J. M., et al. 2001, ApJ, 549, L97
Jenet, F. A., Hobbs, G. B., Lee, K. J., & Manchester, R. N. 2005, ApJ, 625, L123 Suleymanova, S. A., & Shitov, Y. P. 1994, ApJ, 422, 17
Jones, M. L., McLaughlin, M. A., Lam, M. T., et al. 2017, ApJ, 841, 125 Tanenbaum, B. S., Zeissig, G. A., & Drake, F. D. 1968, Science, 160, 760
Karastergiou, A., Johnston, S., Andersson, N., et al. 2015, Advancing Astro- Taylor, J. H. 1992, Phil. Trans. R. Soc. London, Ser. A, 341, 117
physics with the Square Kilometre Array (Macclesfield: SKA Ogranisation), Taylor, J. H., & Huguenin, G. R. 1969, Nature, 221, 816
38 Tiburzi, C. 2018, PASA, 35, e013
Keith, M. J., Coles, W., Shannon, R. M., et al. 2013, MNRAS, 429, 2161 van Haarlem, M. P., Wise, M. W., Gunst, A. W., et al. 2013, A&A, 556, A2
Kerr, M., Coles, W. A., Ward, C. A., et al. 2018, MNRAS, 474, 4637 van Straten, W., & Bailes, M. 2011, PASA, 28, 1
Lam, M. T., Cordes, J. M., Chatterjee, S., et al. 2016, ApJ, 819, 155 van Straten, W., Demorest, P., & Oslowski, S. 2012, Astron. Res. Technol., 9, 237
Lattimer, J. M., & Prakash, M. 2016, Phys. Rep., 621, 127 Verbiest, J. P. W., & Shaifullah, G. 2018, Class. Quant. Grav., 35, 13
Lazarus, P., Karuppusamy, R., Graikou, E., et al. 2016, MNRAS, 458, 868 Verbiest, J. P. W., Weisberg, J. M., Chael, A. A., Lee, K. J., & Lorimer, D. R.
Lazio, T. J. W., Fey, A. L., Dennison, B., et al. 2000, ApJ, 534, 706 2012, ApJ, 755, 39
Lee, K. J., Bassa, C. G., Janssen, G. H., et al. 2014, MNRAS, 441, 2831 Verbiest, J. P. W., Lentati, L., Hobbs, G., et al. 2016, MNRAS, 458, 1267
Lentati, L., Shannon, R. M., Coles, W. A., et al. 2016, MNRAS, 458, 2161 Walker, M. A. 2001, Ap&SS, 278, 149
Liu, S., Pen, U.-L., Macquart, J.-P., Brisken, W., & Deller, A. 2016, MNRAS, Walker, M., & Wardle, M. 1998, ApJ, 498, L125
458, 1289 Will, C. M. 2014, Liv. Rev. Rel., 17, 4
Lorimer, D. R., & Kramer, M. 2005, Handbook of Pulsar Astronomy Yao, J. M., Manchester, R. N., & Wang, N. 2017, ApJ, 835, 29
(Cambridge: Cambridge University Press) You, X.-P., Hobbs, G., Coles, W., et al. 2007, MNRAS, 378, 493

A22, page 11 of 11